{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/dc9b20c2bc55f03a1dd7c18d929a141cc71ef4cbca2e2013c7ffa2642e1fa333.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 0, "uuid": "f627f3ce-06ab-52e0-b919-0882bb98ba87", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.12174165714532137, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f627f3ce-06ab-52e0-b919-0882bb98ba87"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/572a11dae6f7793232f1ecc6d7792779dbd95e0ecbf87877fef7f37fd296deaa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 0, "uuid": "491d8c98-25d3-536b-b156-340c7a47c2d0", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.024486311711370945, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "491d8c98-25d3-536b-b156-340c7a47c2d0"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/efa0d870b83867f8ea856ba39ba809b330fca342415994531f7ecb7529d799ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 14.591754162218422, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "81e94812-2329-5af8-ae49-8167cf27ca32"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 0, "uuid": "81e94812-2329-5af8-ae49-8167cf27ca32", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "81e94812-2329-5af8-ae49-8167cf27ca32"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 25}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/ec61a985467f7c12f4496197b00f64475ab6ac5f5b4e142527235e7ae823a9d0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 25, "Nq": 25}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.008140245918184519, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "361da807-31fa-5a25-8b62-37752b8f3529"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 25, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 0, "uuid": "361da807-31fa-5a25-8b62-37752b8f3529", "validation": {"axes": {"B": 1, "Nq": 25}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "361da807-31fa-5a25-8b62-37752b8f3529"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/2c87c6952f0b539065032448466ba4e943b4e5892d90c8b3a67b5f379a140233.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 0, "uuid": "220902db-7e9b-5d8d-ac0c-4541ef4e4bbd", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.05146070662885904, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "220902db-7e9b-5d8d-ac0c-4541ef4e4bbd"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/c73097f491d3f526d743389cd71d1c69f174db8d2d1030ab6d6fa3fd8a0c606e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 4, "uuid": "ff62c8f2-f87c-586b-b295-7d130dd1a037", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.04754280112683773, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ff62c8f2-f87c-586b-b295-7d130dd1a037"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/398b0e6f0079beab22d824bd2d7732d057a63eab0bf56882656c29eb5bf6dbcd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04182086419314146, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6dbb2b24-0ece-59ae-b07f-05a11948c833"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 4, "uuid": "6dbb2b24-0ece-59ae-b07f-05a11948c833", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "6dbb2b24-0ece-59ae-b07f-05a11948c833"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/e432defa3011150db18920d77084699d442a88e2349d4234518b7993a045960e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 4, "uuid": "59e21fcd-7080-5d70-aa46-c4500984dddc", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.02600828092545271, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "59e21fcd-7080-5d70-aa46-c4500984dddc"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/5c56ec3d96e05c1d64042ed96d3262d2660d28b20330141694069ac6c8af13d9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 4, "uuid": "6ef54aca-2a99-52c6-9c90-3fee81788165", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.0625], "seconds": 0.04886566987261176, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ef54aca-2a99-52c6-9c90-3fee81788165"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/3c7c50a338fb849474ab1ac4d6e12b6892ae5c16050ad027fa3bb2dd41ea297c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 7, "uuid": "8cd405d5-d44c-5568-8ccd-4aae002946f6", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.047110469080507755, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8cd405d5-d44c-5568-8ccd-4aae002946f6"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/29fa773ea9d4948cbb1d8d03edff83594e9f28d4fdc52ee3186bb05e9b3645b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03917623497545719, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aa1790c4-e4c4-53e6-9e27-0dd837883353"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 7, "uuid": "aa1790c4-e4c4-53e6-9e27-0dd837883353", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "aa1790c4-e4c4-53e6-9e27-0dd837883353"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/5c10722b8773a8ec6586cdf8a3722cb5480a25a24a44fcad1784d1797dfa473f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 7, "uuid": "bdc8ab37-9ff6-5499-b47f-64d3a6e67f27", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.02400219812989235, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bdc8ab37-9ff6-5499-b47f-64d3a6e67f27"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/31f9f63df8a657253651b29f9dad25de8a5b00883933c1e9d1006f238cd217c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239230.3971937, "step": 7, "uuid": "1ef85b69-b938-5554-b019-1cd333110897", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.0463019828312099, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1ef85b69-b938-5554-b019-1cd333110897"}, "variant": "rows_4096"}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/0e8a0eb1b0bbd588a87c0110b5b93be119538d4a458790c6de2449c22c941159.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 0, "uuid": "8cd434d9-bb43-53b9-b795-ad7efb0592bc", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.18826853577047586, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8cd434d9-bb43-53b9-b795-ad7efb0592bc"}, "variant": "rows_6144"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/164724b8f736981258f69907c85c883cf083dc5f393900970197af72cc45da0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.10301753412932158, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dc3b6e8c-6201-52c8-9d2e-a190d1562087"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 0, "uuid": "dc3b6e8c-6201-52c8-9d2e-a190d1562087", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "dc3b6e8c-6201-52c8-9d2e-a190d1562087"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 30}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/5182ae66ea455594116b0c41913cb8f9d94a2fedcad7e5a5b4d58b09ca2a88f6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 30, "Nq": 30}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.004042515996843576, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3a45e86c-e517-57d7-92c0-85586056d271"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 30, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 0, "uuid": "3a45e86c-e517-57d7-92c0-85586056d271", "validation": {"axes": {"B": 1, "Nq": 30}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "3a45e86c-e517-57d7-92c0-85586056d271"}}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/f5441516577f6bd6bdd40b2d516564506a857374b1be4dd6dde0bf475febf2cd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 0, "uuid": "bc334c45-f866-5289-891d-4e12672c7e97", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.03487044433131814, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bc334c45-f866-5289-891d-4e12672c7e97"}, "variant": "rows_6144"}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/d01f888681bf35341b734854873d23d2baee4810670db420f469f27f1f244aff.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 0, "uuid": "17f61747-5e4c-5e99-a7c4-f096b147fb60", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.07722086180001497, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "17f61747-5e4c-5e99-a7c4-f096b147fb60"}, "variant": "rows_6144"}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/862b917b34836b4cb5c914ebd438468ab52cde111f8fd09bc92099d58c6d67be.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 4, "uuid": "8bc7ebff-12d1-5ac5-99bb-d18a9cbdb505", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.06412861868739128, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8bc7ebff-12d1-5ac5-99bb-d18a9cbdb505"}, "variant": "rows_6144"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/c59769297a99e59f2b7c707af5d819e62dc0c4efe7320925aead09f72af4d6e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06502548418939114, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2b9361ff-dbd3-598b-95fc-eeec388cc422"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 4, "uuid": "2b9361ff-dbd3-598b-95fc-eeec388cc422", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "2b9361ff-dbd3-598b-95fc-eeec388cc422"}}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/2b48954b969eecec2384645c15d37f3a6bb884cd33ecc45d78fc12488e89dcf4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 4, "uuid": "f8692d25-e816-55f5-acd2-ac13d8e72886", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.031620386987924576, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f8692d25-e816-55f5-acd2-ac13d8e72886"}, "variant": "rows_6144"}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/809d2b82c51d3f7e33329ebf2c4e60e496995095d76f613140e91a9887feae39.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 4, "uuid": "7c1437a2-685d-540a-b50a-04e060aaf3d1", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.0747010437771678, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7c1437a2-685d-540a-b50a-04e060aaf3d1"}, "variant": "rows_6144"}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/5f198676a629599e54864f76f9630326c23ab09bcdd94e89e26445da6898c418.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 7, "uuid": "87ab8f00-5cea-5f74-8f0f-afa1d50d4a9f", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.0760988718830049, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "87ab8f00-5cea-5f74-8f0f-afa1d50d4a9f"}, "variant": "rows_6144"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/a5bfe0abdc234ffa77ca9d6fd583bfa6de4e7f6f921a4aa6aef800ac3e17c910.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.06466197408735752, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0d5a8062-d653-5f8d-89e7-c92004279f8f"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 7, "uuid": "0d5a8062-d653-5f8d-89e7-c92004279f8f", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "0d5a8062-d653-5f8d-89e7-c92004279f8f"}}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/ee2964bf2018a2348bd2142f30f108e23d5a838b247cb26e7f86799ba00910ff.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 7, "uuid": "eb7c6323-04a7-58c5-8478-c695c63634f5", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.030148770194500685, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "eb7c6323-04a7-58c5-8478-c695c63634f5"}, "variant": "rows_6144"}
{"axes": {"M": 6144}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/b43cd19bb9d3d07d9314264664cce99e0c13c5a7e5a892ca9531c5a4fa5e5cfc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239239.1825383, "step": 7, "uuid": "37051054-7c40-5ab2-8960-1bef5fe3fc57", "validation": {"axes": {"M": 6144}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.07489119237288833, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "37051054-7c40-5ab2-8960-1bef5fe3fc57"}, "variant": "rows_6144"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/dc9b20c2bc55f03a1dd7c18d929a141cc71ef4cbca2e2013c7ffa2642e1fa333.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "8dec516f-2b15-55a8-ab77-40cae911fa48", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.042013777419924736, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8dec516f-2b15-55a8-ab77-40cae911fa48"}, "variant": "rows_4096"}
{"axes": {"M": 4125}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/f3f723c566a4f716349447d7ba8bbbf94118c0c4799385d5bc519ec2b551678d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "8a5941ee-6c6f-5312-a860-d0b26263e260", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.02451479807496071, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8a5941ee-6c6f-5312-a860-d0b26263e260"}, "variant": "rows_4125"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/133f657a23877ad38fb6e1a13cfe79fca69d3c0b0a42041db5e61af4de8100c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06613255385309458, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "322fac48-fa66-5a79-aa96-a99c862d24e2"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "322fac48-fa66-5a79-aa96-a99c862d24e2", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "322fac48-fa66-5a79-aa96-a99c862d24e2"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 4125}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/88d8d74e04b85e837291f3803e67f4fcc0c0f34e2c657e377188c36be21a3217.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4125, "Nq": 4125}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.03569585597142577, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f2a1a8e2-460a-549a-a3d1-bdf74bf42f26"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "f2a1a8e2-460a-549a-a3d1-bdf74bf42f26", "validation": {"axes": {"B": 1, "Nq": 4125}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "f2a1a8e2-460a-549a-a3d1-bdf74bf42f26"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/00b199614f81d11ba2763b59714d61372a4c2add5d193d6e896b191f280974e2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "e8446920-1e2c-5681-872e-2de1f62a1c73", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.023155256174504757, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e8446920-1e2c-5681-872e-2de1f62a1c73"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/6456dd42a050590735972eea6e10edba0fc9b03e040bfda0724a0402340eecf8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "5a287253-9d0d-5c95-9930-90afbe51825a", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.04904062207788229, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5a287253-9d0d-5c95-9930-90afbe51825a"}, "variant": "rows_4096"}
{"axes": {"M": 4125}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/9ac5f664ce8eb9c2a335d65460ff23b7a24ccd094765fd9037d575839ce23fba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "717bab44-772a-580e-bd31-8f449703fca5", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.05703906808048487, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "717bab44-772a-580e-bd31-8f449703fca5"}, "variant": "rows_4125"}
{"axes": {"M": 4125}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/c42272f17f7d4589e136ad00595a27ae911f07e87dc318b8d20353a5016f2a6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 0, "uuid": "f49b336e-43c9-5d0e-a40d-277dc443e7e1", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.05304816411808133, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f49b336e-43c9-5d0e-a40d-277dc443e7e1"}, "variant": "rows_4125"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/6e4dd7f63ee604011aba960400d128e5ed7b7bb21e8e6869be5ee2c10cf44be2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 4, "uuid": "fdaddfc3-abb0-51d1-8c0e-560907ff294b", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.04559852974489331, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fdaddfc3-abb0-51d1-8c0e-560907ff294b"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/76201a82d4c5fc6b8643ae7f6a38c9007047cab851982ac1f6239fb9d4b4f354.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06718758028000593, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "76691d76-05ea-597b-a3d5-147ef371c820"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 4, "uuid": "76691d76-05ea-597b-a3d5-147ef371c820", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "76691d76-05ea-597b-a3d5-147ef371c820"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/51bd9c5c57d67f6e9415a85862a0eff814b53655c3f6ed655113b998da86427f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 4, "uuid": "b2595a6b-c707-5c07-bcd6-07da2b26fb72", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.026999427005648613, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b2595a6b-c707-5c07-bcd6-07da2b26fb72"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/526c870125480519a48c9a638517ada4566b78891bcc40259acd566429c1ee30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 4, "uuid": "09f4ca4f-889a-55a0-ad71-c15c37f2a4fa", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.0625], "seconds": 0.045856614131480455, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "09f4ca4f-889a-55a0-ad71-c15c37f2a4fa"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/15b7d5bdf696f9737135f2b0fb9c070e7b7c314a47f94d42f2b7f2220b463f08.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 7, "uuid": "d307d571-5f4f-5ed0-b244-1e87b6dd034b", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.05973345506936312, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d307d571-5f4f-5ed0-b244-1e87b6dd034b"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/7ed73de575477d876ee2981ece7acd672789ba6fb09b3b83b214de5a92cb7ed8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06704403925687075, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "db82b24b-937b-5711-9558-e521b593d474"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 7, "uuid": "db82b24b-937b-5711-9558-e521b593d474", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "db82b24b-937b-5711-9558-e521b593d474"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/934b83cf86460e9b9742b93a6bb81dac6357bd596b3ccc7dc262b8b997166892.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 7, "uuid": "9698844e-f86b-591b-a7c1-fcafc8b838fe", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.026603277772665024, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9698844e-f86b-591b-a7c1-fcafc8b838fe"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/c931b56df8f13f62bc06c41dd9e86900fd4976b4045c68fe113847a21699d288.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239246.744593, "step": 7, "uuid": "4822d7f6-4e76-5a5b-9e53-149df4f45997", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.05827482184395194, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4822d7f6-4e76-5a5b-9e53-149df4f45997"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/dc9b20c2bc55f03a1dd7c18d929a141cc71ef4cbca2e2013c7ffa2642e1fa333.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 0, "uuid": "e42dfaea-65b5-5490-867f-7e1dc340d154", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.03990104701370001, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e42dfaea-65b5-5490-867f-7e1dc340d154"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/10717ae953e6012412024c6acb8b0875873bb08ac3b6ffd0754feb989dd839ee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04190012393519282, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b1e89f95-4b54-5116-9a57-3c29fb4cb095"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 0, "uuid": "b1e89f95-4b54-5116-9a57-3c29fb4cb095", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "b1e89f95-4b54-5116-9a57-3c29fb4cb095"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 32}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/16b2415f217ef91a9a67466dc67b4f5cf097d249377e14a464187e8006435bb2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 32, "Nq": 32}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.003915837965905666, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d90f0f73-d020-5734-8dd7-710a91a624a7"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 32, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 0, "uuid": "d90f0f73-d020-5734-8dd7-710a91a624a7", "validation": {"axes": {"B": 1, "Nq": 32}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "d90f0f73-d020-5734-8dd7-710a91a624a7"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/92fe8c873d3a111e4563dbce2879fd6db88a4b76f03ca3534d44990d8b28f2af.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 0, "uuid": "77d0fee5-a7ee-5516-ad16-68c0193e9fa2", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.024482560344040394, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "77d0fee5-a7ee-5516-ad16-68c0193e9fa2"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/2f7dcef704d0c2fdfcca395b1fe0a176c03e42955129ebb86e6723dfe76e646b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 0, "uuid": "c8c4d9a7-ef98-53ad-8194-887bdb36a542", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.058263967745006084, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c8c4d9a7-ef98-53ad-8194-887bdb36a542"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/5c0f576a693acfced57412c7695e5a49841a899f035f24edc61c2d0696aee7a3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 4, "uuid": "05211974-8e1d-5b1a-8bad-b214cee70353", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.0601740600541234, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "05211974-8e1d-5b1a-8bad-b214cee70353"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/0113132572dd13083893beeb5b5b233cd47fe3d90a64624d717e848ca60f10da.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04426139825955033, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9bdf7136-45be-561a-9076-36477832e84d"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 4, "uuid": "9bdf7136-45be-561a-9076-36477832e84d", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "9bdf7136-45be-561a-9076-36477832e84d"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/de7120ff63282d8cd61da8fa938b9b21402af820bd1b03277d666220a548cf35.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 4, "uuid": "b6a67dd8-dc4d-5060-9d3b-7568993299ba", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.026707913260906935, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b6a67dd8-dc4d-5060-9d3b-7568993299ba"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/f238a5250b37befcbc6e190fdeccbc6628166ecccf2c5bba5914dba4758f168e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 4, "uuid": "22680fff-4c0c-54e1-9bba-ccdac73a7f5f", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.0625], "seconds": 0.061376822646707296, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "22680fff-4c0c-54e1-9bba-ccdac73a7f5f"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/10421991d48239526ec8f05a46c718c16d90cacd498a0ffb4f4c60c1c3d61f3f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "torch.mm", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 7, "uuid": "6382267e-43b7-5d51-a475-3e153db4dd4c", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.05842510890215635, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6382267e-43b7-5d51-a475-3e153db4dd4c"}, "variant": "rows_4096"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 4, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/e1530ed27fbd9f4871ae26741dcb8a00be93a96ab263ece6a72bb327ec06e2c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.044299577828496695, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cce65778-0b96-5f2a-8a7b-16a26e8a2425"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 7, "uuid": "cce65778-0b96-5f2a-8a7b-16a26e8a2425", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "cce65778-0b96-5f2a-8a7b-16a26e8a2425"}}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/18a22a7fae925dc56b2b3c1c43f379196fbfc23e7a31766ef0c7958e4f74b322.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_out.0", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 7, "uuid": "2a724567-1566-5652-ad8c-2546de654e3c", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.026511874049901962, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2a724567-1566-5652-ad8c-2546de654e3c"}, "variant": "rows_4096"}
{"axes": {"M": 4096}, "capture_steps": [0, 4, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/473a0e2997cbabadc85352895a553f7f9a824a67b08132b7309acdd1c14167d2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 1, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 588, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 1, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha", "requested_outputs": 1, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b1-r1/results/images", "--input-save-path", "/task/raw/image-b1-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239254.438051, "step": 7, "uuid": "ed5569af-287c-5fd3-b99a-3f6fd87e2a33", "validation": {"axes": {"M": 4096}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.06076050875708461, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ed5569af-287c-5fd3-b99a-3f6fd87e2a33"}, "variant": "rows_4096"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/b3b99ef0fc63956f514c1d2a3a83e93eeed5b53defdb7dce3c8c7e309b6d6af2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 0, "uuid": "08b43bd3-126f-54fd-bedc-24656a3f3ba8", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.030111619271337986, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "08b43bd3-126f-54fd-bedc-24656a3f3ba8"}, "variant": "rows_8192"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/efa0d870b83867f8ea856ba39ba809b330fca342415994531f7ecb7529d799ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 20.078923034016043, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0fded79d-d450-5a30-a52a-054d2aba0b5e"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 0, "uuid": "0fded79d-d450-5a30-a52a-054d2aba0b5e", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "0fded79d-d450-5a30-a52a-054d2aba0b5e"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 25}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/ec61a985467f7c12f4496197b00f64475ab6ac5f5b4e142527235e7ae823a9d0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 25, "Nq": 25}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.006427989806979895, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8d248f2d-1847-5526-a70d-1382a83477f8"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 25, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 0, "uuid": "8d248f2d-1847-5526-a70d-1382a83477f8", "validation": {"axes": {"B": 1, "Nq": 25}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "8d248f2d-1847-5526-a70d-1382a83477f8"}}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/224e353933f43af734ec74834219cd0d532f8a4c2a5161f1bd6cb684e9f0c653.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 0, "uuid": "b571dcbc-df9a-515c-90ac-93ba5e03a0dd", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.24693757621571422, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b571dcbc-df9a-515c-90ac-93ba5e03a0dd"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/b9fa10811c2bd2c54d53afbac8b7c887ca75deec828a11123642283d5266b98e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 0, "uuid": "a870fe23-b82a-5273-bdfe-7d289bbe9537", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.05861895019188523, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a870fe23-b82a-5273-bdfe-7d289bbe9537"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/e9a32d3ed18cffbada4d9c8f2ecca15565f466ed74fbacaa98f9d5b72bf7137a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 7, "uuid": "90021c8c-9d0c-55ea-857e-e12ad4fe7979", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.027585350442677736, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "90021c8c-9d0c-55ea-857e-e12ad4fe7979"}, "variant": "rows_8192"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/29fa773ea9d4948cbb1d8d03edff83594e9f28d4fdc52ee3186bb05e9b3645b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03551399428397417, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7eb997b9-50e5-59e0-8a3e-15cf20c6e085"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 7, "uuid": "7eb997b9-50e5-59e0-8a3e-15cf20c6e085", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "7eb997b9-50e5-59e0-8a3e-15cf20c6e085"}}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/58968e9db4efaafc3ee4bba457398afca703cd0aa049f476f437399dfddf4732.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 7, "uuid": "b9572ad2-03a8-5830-8ded-11cb591ccd78", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.05876733176410198, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b9572ad2-03a8-5830-8ded-11cb591ccd78"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/1ef5e531605f82e019d314bcd63a1457da4f407923461967801d89a06bc92e5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239366.2158682, "step": 7, "uuid": "d26eaeed-136a-5399-999a-baf0a72a8818", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.057101557962596416, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d26eaeed-136a-5399-999a-baf0a72a8818"}, "variant": "rows_8192"}
{"axes": {"M": 12288}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/ff6b72139b04fa9130144a63bab15c6ff519505fd4d6f326cc23ee482a7ec552.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 0, "uuid": "67b173dc-ff53-562e-be4c-8e5d7b603f06", "validation": {"axes": {"M": 12288}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.04194791987538338, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "67b173dc-ff53-562e-be4c-8e5d7b603f06"}, "variant": "rows_12288"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/164724b8f736981258f69907c85c883cf083dc5f393900970197af72cc45da0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.08632527804002166, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b195ce43-7a00-5c7e-8570-4081ae872fdb"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 0, "uuid": "b195ce43-7a00-5c7e-8570-4081ae872fdb", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "b195ce43-7a00-5c7e-8570-4081ae872fdb"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 30}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/5182ae66ea455594116b0c41913cb8f9d94a2fedcad7e5a5b4d58b09ca2a88f6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 30, "Nq": 30}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.0024891369976103306, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "92594a4b-163e-553b-b0e6-ee5932121fe0"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 30, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 0, "uuid": "92594a4b-163e-553b-b0e6-ee5932121fe0", "validation": {"axes": {"B": 1, "Nq": 30}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "92594a4b-163e-553b-b0e6-ee5932121fe0"}}
{"axes": {"M": 12288}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/e8cd7bb7a15651b159c4649f4ab20dac14a78ba1ab3ce7494ac9f2b1948d4580.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 0, "uuid": "d89ba855-bc0f-52a5-9825-2f44455164ea", "validation": {"axes": {"M": 12288}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.3068883470259607, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d89ba855-bc0f-52a5-9825-2f44455164ea"}, "variant": "rows_12288"}
{"axes": {"M": 12288}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/805484a6218332a054439c789defb50bf2195a889cbd859ffab2e6bbe8ed2550.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 0, "uuid": "f53807e9-dd75-5cca-9c00-8cc6f48f3a3a", "validation": {"axes": {"M": 12288}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.08012292999774218, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f53807e9-dd75-5cca-9c00-8cc6f48f3a3a"}, "variant": "rows_12288"}
{"axes": {"M": 12288}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/c8680a8edeb1c05535a2e0c3911596054a84f108b79ebd553df6d9d0878a4e31.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 7, "uuid": "0be5412a-a24b-59e7-8cc6-ec6de9b0e6a1", "validation": {"axes": {"M": 12288}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.03814372327178717, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0be5412a-a24b-59e7-8cc6-ec6de9b0e6a1"}, "variant": "rows_12288"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/a5bfe0abdc234ffa77ca9d6fd583bfa6de4e7f6f921a4aa6aef800ac3e17c910.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.05036327103152871, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9a5607a1-3d45-5614-9c28-556a7235eaf5"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 7, "uuid": "9a5607a1-3d45-5614-9c28-556a7235eaf5", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "9a5607a1-3d45-5614-9c28-556a7235eaf5"}}
{"axes": {"M": 12288}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/556b48d2e95940c9e20e93bd8cbcfa92e97c40b958bec2affc7e5c540af69a6a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 7, "uuid": "1e9d5954-f8dc-5e61-b3e3-4a7adad70ad6", "validation": {"axes": {"M": 12288}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07530533708631992, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1e9d5954-f8dc-5e61-b3e3-4a7adad70ad6"}, "variant": "rows_12288"}
{"axes": {"M": 12288}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/515b571e521810128a185851cdedb170a79903d6ad7dcdd4e896540dfc986b7e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239383.7062342, "step": 7, "uuid": "8886990a-4593-57b7-bd60-9c00e411ab13", "validation": {"axes": {"M": 12288}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.07717035384848714, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8886990a-4593-57b7-bd60-9c00e411ab13"}, "variant": "rows_12288"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/b3b99ef0fc63956f514c1d2a3a83e93eeed5b53defdb7dce3c8c7e309b6d6af2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "b36852c0-2e4d-5b24-92ce-2c46d5fc8c97", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.028228317853063345, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b36852c0-2e4d-5b24-92ce-2c46d5fc8c97"}, "variant": "rows_8192"}
{"axes": {"M": 4125}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/f3f723c566a4f716349447d7ba8bbbf94118c0c4799385d5bc519ec2b551678d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "30a38acb-ba44-550d-b1ff-48921f60e605", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.01648318814113736, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "30a38acb-ba44-550d-b1ff-48921f60e605"}, "variant": "rows_4125"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/133f657a23877ad38fb6e1a13cfe79fca69d3c0b0a42041db5e61af4de8100c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0512183029204607, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8ac22de4-c54c-594a-9759-904f1c30f6f7"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "8ac22de4-c54c-594a-9759-904f1c30f6f7", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "8ac22de4-c54c-594a-9759-904f1c30f6f7"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 4125}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/88d8d74e04b85e837291f3803e67f4fcc0c0f34e2c657e377188c36be21a3217.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4125, "Nq": 4125}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.028184604365378618, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a1487fbc-a4dc-5724-9bd9-0faf4c062b6b"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "a1487fbc-a4dc-5724-9bd9-0faf4c062b6b", "validation": {"axes": {"B": 1, "Nq": 4125}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "a1487fbc-a4dc-5724-9bd9-0faf4c062b6b"}}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/0ef04f05a00478a89c1b4b0188148efef283f0edd09d456eb27ec1b535c0a063.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "15ceda36-6fd7-5ea2-9889-1d88eb4495ab", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.055546706076711416, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "15ceda36-6fd7-5ea2-9889-1d88eb4495ab"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/bca455937434011b259aa80f80861a54352a79daa077a940ed3cdf50ea0dd0d1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "857a2e8b-b80a-5f8c-825d-ff2fe2edbfa0", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.05560618685558438, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "857a2e8b-b80a-5f8c-825d-ff2fe2edbfa0"}, "variant": "rows_8192"}
{"axes": {"M": 4125}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/9ac5f664ce8eb9c2a335d65460ff23b7a24ccd094765fd9037d575839ce23fba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "e2d459ae-d79f-575f-aa23-3fdc5824cdda", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03876070398837328, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e2d459ae-d79f-575f-aa23-3fdc5824cdda"}, "variant": "rows_4125"}
{"axes": {"M": 4125}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/c42272f17f7d4589e136ad00595a27ae911f07e87dc318b8d20353a5016f2a6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 0, "uuid": "d2d098b8-07b6-5ca5-8a79-9f6f7e8f93cd", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.040546406991779804, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d2d098b8-07b6-5ca5-8a79-9f6f7e8f93cd"}, "variant": "rows_4125"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/7f4080a6c5555bf12e9aaf6757822efc462b9067ead0da2ed669c0e809e545e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 7, "uuid": "0b1369ce-f541-5687-bff0-28b54246e54d", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.030160407070070505, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0b1369ce-f541-5687-bff0-28b54246e54d"}, "variant": "rows_8192"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/7ed73de575477d876ee2981ece7acd672789ba6fb09b3b83b214de5a92cb7ed8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.051746082957834005, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "85c2aa46-93b8-5725-ba09-29913e05e1cd"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 7, "uuid": "85c2aa46-93b8-5725-ba09-29913e05e1cd", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "85c2aa46-93b8-5725-ba09-29913e05e1cd"}}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/2853a80fc4317a86f06b98b502e7ab616e225fb12947dbfd1a21463a2a8bbac1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 7, "uuid": "cb09899b-2462-5f7b-a3df-1b8bf624f35c", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.059363945852965117, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cb09899b-2462-5f7b-a3df-1b8bf624f35c"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/765cfebcce81c1dfa6b40b4e39376d0997ba89bfc006133c7a22d439ecbef9ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239393.0137348, "step": 7, "uuid": "acec1a44-43f2-5f4b-b13f-237d2e82cf96", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.058005810249596834, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "acec1a44-43f2-5f4b-b13f-237d2e82cf96"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/b3b99ef0fc63956f514c1d2a3a83e93eeed5b53defdb7dce3c8c7e309b6d6af2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 0, "uuid": "b4625364-7ec2-5ca7-9997-48c6aac77557", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.02641299506649375, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b4625364-7ec2-5ca7-9997-48c6aac77557"}, "variant": "rows_8192"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/10717ae953e6012412024c6acb8b0875873bb08ac3b6ffd0754feb989dd839ee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03146282723173499, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aa86e0ed-fe05-5873-998c-df6ff7963aae"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 0, "uuid": "aa86e0ed-fe05-5873-998c-df6ff7963aae", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "aa86e0ed-fe05-5873-998c-df6ff7963aae"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 32}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/16b2415f217ef91a9a67466dc67b4f5cf097d249377e14a464187e8006435bb2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 32, "Nq": 32}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.00240721320733428, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b8441206-1742-5b50-a686-e0d86149ca28"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 32, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 0, "uuid": "b8441206-1742-5b50-a686-e0d86149ca28", "validation": {"axes": {"B": 1, "Nq": 32}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "b8441206-1742-5b50-a686-e0d86149ca28"}}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/45bc577fdd6156c87de4860bb22460fa8a8f021f385f9baf8b47b572d5d2c3f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 0, "uuid": "6fc5035f-9974-5364-a58c-0c21a341f149", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.05917341588065028, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6fc5035f-9974-5364-a58c-0c21a341f149"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/49f73531422709582089fb5916de279b062ce9a9cc79d873acc234d20b0f542a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 0, "uuid": "910fa072-ce50-558a-a4d7-f6b68ab8f7f3", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.058030262123793364, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "910fa072-ce50-558a-a4d7-f6b68ab8f7f3"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/fd1449aad6fde5b85393e2a05b70531c691604ad0ed29800eca0de48a9bd3be7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 7, "uuid": "eecc9d19-a7b9-59bb-bcf3-12ee8da5d8f4", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.030127210076898336, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "eecc9d19-a7b9-59bb-bcf3-12ee8da5d8f4"}, "variant": "rows_8192"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/e1530ed27fbd9f4871ae26741dcb8a00be93a96ab263ece6a72bb327ec06e2c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03204311802983284, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "df7932e0-93d3-5be4-92af-f6eba34890a5"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 7, "uuid": "df7932e0-93d3-5be4-92af-f6eba34890a5", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "df7932e0-93d3-5be4-92af-f6eba34890a5"}}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/2587d48e43962811485244f272459d67d7e7ffd396f95c02125b9fd9457aa925.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 7, "uuid": "0882664a-03ac-5037-b167-fa264ffa0639", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.05981408432126045, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0882664a-03ac-5037-b167-fa264ffa0639"}, "variant": "rows_8192"}
{"axes": {"M": 8192}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/df205ade4853b5cd057d8112d25d7a8ea8f3ead9e1b8a431c22457771eedba6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 2, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 583, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 2, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch2", "requested_outputs": 2, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b2-r1/results/images", "--input-save-path", "/task/raw/image-b2-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239400.7528443, "step": 7, "uuid": "d301ee66-c6e8-5668-aacd-b5c4838e6bbe", "validation": {"axes": {"M": 8192}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.05757515178993344, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d301ee66-c6e8-5668-aacd-b5c4838e6bbe"}, "variant": "rows_8192"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/bf1eafcbcc197bf05a07281267e6cbf894cdb701ebdb46b14f39eba6596c16ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 0, "uuid": "1a8fd528-d8a7-533c-be10-6bbb74925548", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.08350890874862671, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a8fd528-d8a7-533c-be10-6bbb74925548"}, "variant": "rows_16384"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/efa0d870b83867f8ea856ba39ba809b330fca342415994531f7ecb7529d799ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 17.520813889801502, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5ad9db13-318b-5717-a1fd-6bafe619bb28"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 0, "uuid": "5ad9db13-318b-5717-a1fd-6bafe619bb28", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "5ad9db13-318b-5717-a1fd-6bafe619bb28"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 25}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/ec61a985467f7c12f4496197b00f64475ab6ac5f5b4e142527235e7ae823a9d0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 25, "Nq": 25}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.007903106976300478, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6f454b8-c97f-582d-a63c-faa18a6f6ad9"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 25, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 0, "uuid": "c6f454b8-c97f-582d-a63c-faa18a6f6ad9", "validation": {"axes": {"B": 1, "Nq": 25}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "c6f454b8-c97f-582d-a63c-faa18a6f6ad9"}}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/fb4921609679476f4798dd212bf8c2664f5010298a19a7c8b31f339899216163.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 0, "uuid": "6c24a9cf-df14-538e-9876-3431874e3701", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.4345134859904647, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6c24a9cf-df14-538e-9876-3431874e3701"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/c9423d86c669bcfe363c735213067aa4e2c7692d87bbbd31b0d5657ef6fb524f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 0, "uuid": "a5375c67-0c90-5d89-9dc9-808e7ec3b6f2", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.11248841416090727, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a5375c67-0c90-5d89-9dc9-808e7ec3b6f2"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/112b591be58c8366bf508e157055684de7853bcd091a062d2ad10fa00b45cfad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 7, "uuid": "4995e0dc-1c17-5f42-9530-b660dce2f09d", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.08605304500088096, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4995e0dc-1c17-5f42-9530-b660dce2f09d"}, "variant": "rows_16384"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 25, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/29fa773ea9d4948cbb1d8d03edff83594e9f28d4fdc52ee3186bb05e9b3645b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4121, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04081157920882106, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "301ea11f-ef1e-5dbf-903c-f53f830b7f49"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 25, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 7, "uuid": "301ea11f-ef1e-5dbf-903c-f53f830b7f49", "validation": {"axes": {"B": 1, "Np": 25, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "301ea11f-ef1e-5dbf-903c-f53f830b7f49"}}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/5f7258afce97231dcfe4e69124dffde02e063b9c2a801460c13620e5e1d10b6c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 7, "uuid": "f8511c9e-edb1-5002-a586-b37dbfb060bc", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.16125830775126815, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f8511c9e-edb1-5002-a586-b37dbfb060bc"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/bc16b5eaf4b699260d2bf850ec26ddaa0762588aedf1721e599aaab5cfd4a65d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A red ceramic teapot on a wooden table beside a window, soft afternoon light.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_square_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239509.878295, "step": 7, "uuid": "f567a1a4-8780-53bc-bd48-6b7d561d8eb8", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.11023131385445595, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f567a1a4-8780-53bc-bd48-6b7d561d8eb8"}, "variant": "rows_16384"}
{"axes": {"M": 24576}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/6d2216c22331e042697018e03e769e31d54511e186e634c64997144158095394.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 0, "uuid": "54ea04a7-b4b5-5279-bed3-34f4ae1cf45f", "validation": {"axes": {"M": 24576}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.07319077430292964, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "54ea04a7-b4b5-5279-bed3-34f4ae1cf45f"}, "variant": "rows_24576"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/164724b8f736981258f69907c85c883cf083dc5f393900970197af72cc45da0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.10544117726385593, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "62e5d87f-9f4f-52c8-96b4-5c97c8932028"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 0, "uuid": "62e5d87f-9f4f-52c8-96b4-5c97c8932028", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "62e5d87f-9f4f-52c8-96b4-5c97c8932028"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 30}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/5182ae66ea455594116b0c41913cb8f9d94a2fedcad7e5a5b4d58b09ca2a88f6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 30, "Nq": 30}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.003906491212546825, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5f8f619a-7132-5f12-ac6c-ef19bcb3fbd0"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 30, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 0, "uuid": "5f8f619a-7132-5f12-ac6c-ef19bcb3fbd0", "validation": {"axes": {"B": 1, "Nq": 30}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "5f8f619a-7132-5f12-ac6c-ef19bcb3fbd0"}}
{"axes": {"M": 24576}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/578bf17c5693607e5cbffd8533249b100ee1a1b686c3b750009ab41f6dc8a110.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 0, "uuid": "8c811c7c-db79-50a4-a48d-c1e483e9418b", "validation": {"axes": {"M": 24576}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.7522421278990805, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8c811c7c-db79-50a4-a48d-c1e483e9418b"}, "variant": "rows_24576"}
{"axes": {"M": 24576}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/72927b582e12d71d289c2ceccab3f36c6d7c71123b7dabc1407380e5e40d47fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 0, "uuid": "89b32f3e-3660-5591-8885-48dffb786922", "validation": {"axes": {"M": 24576}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.16783020878210664, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "89b32f3e-3660-5591-8885-48dffb786922"}, "variant": "rows_24576"}
{"axes": {"M": 24576}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/774f7b9a78d8bf988776e979aa8aa5b4c63544a39275df5ee0713d2d6ffdd217.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 7, "uuid": "4d48e098-93c0-5f3c-867b-cb99441d7db7", "validation": {"axes": {"M": 24576}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.10091953771188855, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4d48e098-93c0-5f3c-867b-cb99441d7db7"}, "variant": "rows_24576"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 30, "Nq": 6144}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/a5bfe0abdc234ffa77ca9d6fd583bfa6de4e7f6f921a4aa6aef800ac3e17c910.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 6174, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.06686250399798155, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5853c35f-a657-5720-8165-27f421255063"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 30, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 7, "uuid": "5853c35f-a657-5720-8165-27f421255063", "validation": {"axes": {"B": 1, "Np": 30, "Nq": 6144}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "5853c35f-a657-5720-8165-27f421255063"}}
{"axes": {"M": 24576}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/b995b47f035c0878afab3db51944cc4a5336bc7f1be525a9b28427542b9d1a32.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 7, "uuid": "304cdacf-4248-51bd-b9b7-1f2bdcde0405", "validation": {"axes": {"M": 24576}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.1714622601866722, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "304cdacf-4248-51bd-b9b7-1f2bdcde0405"}, "variant": "rows_24576"}
{"axes": {"M": 24576}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/7837393a11af2cce97f6b95056b11658016dc12a3f2b7eeed6efa030b5f83cc2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A quiet lakeside village at sunrise, reflections in the water, mountains behind the houses, wide landscape photograph.", "response_format": "b64_json", "seed": 42, "size": "1536x1024"}, "request_id": "qwen_image21_landscape_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239569.4210482, "step": 7, "uuid": "3589b4f7-d841-5091-87f0-84fc90925259", "validation": {"axes": {"M": 24576}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.23707909928634763, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3589b4f7-d841-5091-87f0-84fc90925259"}, "variant": "rows_24576"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/bf1eafcbcc197bf05a07281267e6cbf894cdb701ebdb46b14f39eba6596c16ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "0de79c88-bf5d-53f0-9f99-77f49766480b", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.07636397518217564, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0de79c88-bf5d-53f0-9f99-77f49766480b"}, "variant": "rows_16384"}
{"axes": {"M": 4125}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/f3f723c566a4f716349447d7ba8bbbf94118c0c4799385d5bc519ec2b551678d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "c476dbd5-9a1f-56b8-b4fe-4dd239d9a8f2", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.032974231988191605, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c476dbd5-9a1f-56b8-b4fe-4dd239d9a8f2"}, "variant": "rows_4125"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/133f657a23877ad38fb6e1a13cfe79fca69d3c0b0a42041db5e61af4de8100c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.07032340299338102, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "985e86ef-7f75-5ae0-a11f-2b16728e0a16"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "985e86ef-7f75-5ae0-a11f-2b16728e0a16", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "985e86ef-7f75-5ae0-a11f-2b16728e0a16"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 4125}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/88d8d74e04b85e837291f3803e67f4fcc0c0f34e2c657e377188c36be21a3217.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4125, "Nq": 4125}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.03707037400454283, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2f90dd5b-3c00-5a2c-8767-9a38ea6897ac"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "2f90dd5b-3c00-5a2c-8767-9a38ea6897ac", "validation": {"axes": {"B": 1, "Nq": 4125}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "2f90dd5b-3c00-5a2c-8767-9a38ea6897ac"}}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/3d48305d939f178cd8b60aae1986f9fd0004e2649a3b0900ebc4481683cf74bf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "f662784e-03cd-58da-8401-07b85a5b5e2e", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.1718691810965538, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f662784e-03cd-58da-8401-07b85a5b5e2e"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/467ade334037b842da3b6d1aca2336af14fa8934db3493d60e990f77f75a755c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "8ce49068-3af8-5772-90d3-913923d2cea7", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.12243430502712727, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8ce49068-3af8-5772-90d3-913923d2cea7"}, "variant": "rows_16384"}
{"axes": {"M": 4125}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/9ac5f664ce8eb9c2a335d65460ff23b7a24ccd094765fd9037d575839ce23fba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "5319e423-c7fa-5af2-a356-948cb73ad345", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.05945582827553153, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5319e423-c7fa-5af2-a356-948cb73ad345"}, "variant": "rows_4125"}
{"axes": {"M": 4125}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/c42272f17f7d4589e136ad00595a27ae911f07e87dc318b8d20353a5016f2a6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 0, "uuid": "a5a627dc-d290-5391-a62e-6bad43f1751a", "validation": {"axes": {"M": 4125}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.054225822910666466, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a5a627dc-d290-5391-a62e-6bad43f1751a"}, "variant": "rows_4125"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/6b26c4b5ae9a8acbdeaace0403da7cd75eae5de4013f849cd77dbd531ef186f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 7, "uuid": "c28c370a-48bd-5f55-8b1b-4b915ecbce09", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.07907597534358501, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c28c370a-48bd-5f55-8b1b-4b915ecbce09"}, "variant": "rows_16384"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 4125, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/7ed73de575477d876ee2981ece7acd672789ba6fb09b3b83b214de5a92cb7ed8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 8221, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06200115615502, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "881d21a7-fe05-55e3-8a3f-fbed1a05e62a"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 8, false], [8, 4104, true], [4104, 4125, false]], "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 7, "uuid": "881d21a7-fe05-55e3-8a3f-fbed1a05e62a", "validation": {"axes": {"B": 1, "Np": 4125, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "881d21a7-fe05-55e3-8a3f-fbed1a05e62a"}}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/e494f5abd9f8586b4bd24f473c3cfd45f2ec4b13eadc2e8946cc2a72bd4860db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 7, "uuid": "def17f80-03dc-5b52-bf7f-2ba88e69d105", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.11522956704720855, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "def17f80-03dc-5b52-bf7f-2ba88e69d105"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/3a0b17e7438e60e0a47fd576c2ce85c3d44c26f667f8ee1e0fd4e4f56b8cc5cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "Change the teapot to cobalt blue while preserving the composition and lighting.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_edit_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239587.3535016, "step": 7, "uuid": "b630cff7-4855-58a6-b355-47e6dc0d15ea", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.12029199302196503, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b630cff7-4855-58a6-b355-47e6dc0d15ea"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/bf1eafcbcc197bf05a07281267e6cbf894cdb701ebdb46b14f39eba6596c16ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 0, "uuid": "4d4a7efc-7dec-550c-b41c-ea5c87568f7a", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [0.5], "seconds": 0.07532536098733544, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4d4a7efc-7dec-550c-b41c-ea5c87568f7a"}, "variant": "rows_16384"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/10717ae953e6012412024c6acb8b0875873bb08ac3b6ffd0754feb989dd839ee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03661294002085924, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f9129b9a-80f1-5c9e-9589-1697a7b4738d"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 0, "uuid": "f9129b9a-80f1-5c9e-9589-1697a7b4738d", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "f9129b9a-80f1-5c9e-9589-1697a7b4738d"}}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nq": 32}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128_prefix0", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128_prefix0/16b2415f217ef91a9a67466dc67b4f5cf097d249377e14a464187e8006435bb2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 32, "Nq": 32}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.002510282676666975, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1c24ae50-e45d-5786-a2d9-36809250cba9"}, "removed_capture_only_inputs": ["K_prefix", "V_prefix", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn", "operation": "QwenImage21Attention.local_attn (all prefix segments)", "prefix_segments": [[0, 32, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 0, "uuid": "1c24ae50-e45d-5786-a2d9-36809250cba9", "validation": {"axes": {"B": 1, "Nq": 32}, "definition": "prefix_cached_block_causal_h32_d128_prefix0", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "1c24ae50-e45d-5786-a2d9-36809250cba9"}}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/aa09b35816f4ec0199d1f339bc9b06750db3dc167f098a51101f6d94b83b1fea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 0, "uuid": "f47fbafc-f0a9-518a-ad27-9ed12d8832c1", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.11258525215089321, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f47fbafc-f0a9-518a-ad27-9ed12d8832c1"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/d57d927e98006f51e8fb72bbbd5ff24d92dadc4be43b5faf0646271bbc8b3a2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 0, "uuid": "8344cb3d-646a-5983-955d-aa29071bc411", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.125], "seconds": 0.1201779879629612, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8344cb3d-646a-5983-955d-aa29071bc411"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k4096/5c64aa0bb4886fb265252031d7f183da548b6dc1d4196cb1e5e2dcccf5134689.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.to_q", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 7, "uuid": "8324365f-527d-5871-ba3c-5bc0672278aa", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k4096", "device": "cuda", "max_absolute_errors": [1.0], "seconds": 0.08119806088507175, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8324365f-527d-5871-ba3c-5bc0672278aa"}, "variant": "rows_16384"}
{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Np": 32, "Nq": 4096}, "capture_steps": [0, 7], "definition": "prefix_cached_block_causal_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/prefix_cached_block_causal_h32_d128/e1530ed27fbd9f4871ae26741dcb8a00be93a96ab263ece6a72bb327ec06e2c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nall": 4128, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03464225912466645, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c8e32486-c5c6-5f42-bfbe-8d1f7c0a66d5"}, "removed_capture_only_inputs": ["image_ids", "key_valid", "sm_scale"], "replay_date": "2026-10-19"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.attn.target_attn", "operation": "QwenImage21Attention.target_attn", "prefix_segments": [[0, 32, false]], "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 7, "uuid": "c8e32486-c5c6-5f42-bfbe-8d1f7c0a66d5", "validation": {"axes": {"B": 1, "Np": 32, "Nq": 4096}, "definition": "prefix_cached_block_causal_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "c8e32486-c5c6-5f42-bfbe-8d1f7c0a66d5"}}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n12288_k4096", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n12288_k4096/52242c387e9b0142be33380829460bac47428c61cd3efeeacced4d883086cd0c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.gate_layer", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 7, "uuid": "3f471772-09e1-5090-a367-35799e617731", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n12288_k4096", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.1156368306837976, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3f471772-09e1-5090-a367-35799e617731"}, "variant": "rows_16384"}
{"axes": {"M": 16384}, "capture_steps": [0, 7], "definition": "dense_gemm_bf16_nt_n4096_k12288", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/dense_gemm_bf16_nt_n4096_k12288/cdfecb022e906114110dbb7e9699d93328a4f54d2c3ee44cd5285524119c66dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Qwen/Qwen-Image-2.1-Turbo", "model_batch_size": 4, "model_revision": "fc13ba1986fcb905854dc8de2f52eee395f49b5b", "module_path": "transformer_blocks.0.img_mlp.out", "operation": "torch.nn.functional.linear", "process_id": 581, "request": {"guidance_scale": 1.0, "model": "Qwen-Image-2.1", "n": 4, "num_inference_steps": 8, "output_format": "png", "prompt": "A single orange cat sitting, full body, isolated on a transparent background with an alpha channel. No floor or shadow.", "response_format": "b64_json", "seed": 42, "size": "1024x1024"}, "request_id": "qwen_image21_alpha_batch4", "requested_outputs": 4, "sampling_adapter": "Use the unmodified Turbo schedule stored in model_index.json.", "sampling_sigmas": [1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568], "server_command": ["sglang", "serve", "--model-path", "/task/models/qwen-image21", "--model-id", "Qwen-Image-2.1", "--host", "127.0.0.1", "--port", "30104", "--scheduler-port", "30124", "--master-port", "30125", "--performance-mode", "speed", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--output-path", "/task/raw/image-b4-r1/results/images", "--input-save-path", "/task/raw/image-b4-r1/results/inputs"], "source": "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo", "started_at": 1792239601.4219952, "step": 7, "uuid": "d9ff2a4f-4445-5c06-92ec-f19bd4c39911", "validation": {"axes": {"M": 16384}, "definition": "dense_gemm_bf16_nt_n4096_k12288", "device": "cuda", "max_absolute_errors": [0.25], "seconds": 0.10815242305397987, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d9ff2a4f-4445-5c06-92ec-f19bd4c39911"}, "variant": "rows_16384"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/2ac2a1d1e80676c131af9c5b87b012ab989b4d17087129af1e3efc364a8aee91.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.20056561101228, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c7ba0aad-2212-5bb2-8896-64d663665b59"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "c7ba0aad-2212-5bb2-8896-64d663665b59", "validation": {"axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "c7ba0aad-2212-5bb2-8896-64d663665b59"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/3f45356cd73e6d1045cf350d684ebac058e60d9b764ff34e85d9a092c6cf32f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0001220703125], "seconds": 4.497790469322354, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c884d64e-9420-501c-b198-7807cdda2443"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "c884d64e-9420-501c-b198-7807cdda2443", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0001220703125], "status": "passed", "uuid": "c884d64e-9420-501c-b198-7807cdda2443"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/793861d1494df006ca325c4276e8aade9eaa15a9dbd82dd3d33d74ffcb236316.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.09004547726362944, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "40268ed4-06c5-5827-81ef-1c9ee0ee1538"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "40268ed4-06c5-5827-81ef-1c9ee0ee1538", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "40268ed4-06c5-5827-81ef-1c9ee0ee1538"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/21e5aa860b4bb207de53f96fc4d00b1d318c2b6faaaed12193f620ebbefc69ce.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.009442159906029701, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2f9ae066-bf34-53c8-bb42-abe64c739e10"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "2f9ae066-bf34-53c8-bb42-abe64c739e10", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "2f9ae066-bf34-53c8-bb42-abe64c739e10"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/bfe6c981523624929104cfa3e2b4d5f4f80000012e0388efdc1efa04224de998.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.017903539817780256, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bfe2c0d7-2dd2-582f-bd46-af8b3b9dc217"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "bfe2c0d7-2dd2-582f-bd46-af8b3b9dc217", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "bfe2c0d7-2dd2-582f-bd46-af8b3b9dc217"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f432c0d559af1a91ec604bb9dfcd33f18132ad81db7255a6401ffd9d3cb789a8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.009230131283402443, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9182253a-2949-52dd-8da2-d2ebea4c91fc"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "9182253a-2949-52dd-8da2-d2ebea4c91fc", "validation": {"axes": {"B": 1, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "9182253a-2949-52dd-8da2-d2ebea4c91fc"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/5897e3b872a866c908f982974aa7b80ca7dba00e982f64d0d34802df82562d1b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.05460761534050107, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cb8dd366-a40b-5ea5-a035-80beb2c63a25"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 4, "uuid": "cb8dd366-a40b-5ea5-a035-80beb2c63a25", "validation": {"axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "cb8dd366-a40b-5ea5-a035-80beb2c63a25"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/b4161fe9a82546194c9df7a8fc954778841d917f99a92fae83e6402c6cc4474d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.002057234290987253, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "82ffc151-9721-5a67-b4a1-2fa8c886b84b"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 4, "uuid": "82ffc151-9721-5a67-b4a1-2fa8c886b84b", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "82ffc151-9721-5a67-b4a1-2fa8c886b84b"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/cced123163ad4947c76601df2d55ae2cb5907a701afecbf4cf44e1ea3324a87c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.02799592912197113, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c80509df-cd6d-52cb-b405-6c4eb5efee7f"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 4, "uuid": "c80509df-cd6d-52cb-b405-6c4eb5efee7f", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "c80509df-cd6d-52cb-b405-6c4eb5efee7f"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/1b456c5165e08733548e8b07913faf9f402b3c21632c7ebbc2d38239cb354ab1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.004200323950499296, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d4d64ece-f3a9-5c54-8d94-5c89afa12f64"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 4, "uuid": "d4d64ece-f3a9-5c54-8d94-5c89afa12f64", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00048828125], "status": "passed", "uuid": "d4d64ece-f3a9-5c54-8d94-5c89afa12f64"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f5e02816c9e6c4bda246bd18b24cf9a7362d3a5e66046222f23fd4fe5eb77a2e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.00674821762368083, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "22824146-c2bc-5edb-8ec1-e14d5a6ddfe7"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 4, "uuid": "22824146-c2bc-5edb-8ec1-e14d5a6ddfe7", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "22824146-c2bc-5edb-8ec1-e14d5a6ddfe7"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/d796ec3d2a2ceb01f649054a35b9cbf9f1184f0a3aaa7d5a3d0cfa8c656399ee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.006750067695975304, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e7cff504-efb7-5e51-83ca-394ac66f04fb"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 4, "uuid": "e7cff504-efb7-5e51-83ca-394ac66f04fb", "validation": {"axes": {"B": 1, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0], "status": "passed", "uuid": "e7cff504-efb7-5e51-83ca-394ac66f04fb"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/982cf3dcd23cf10f12351e3df1be784376463253d56dcea27f9516743a403766.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.05405203020200133, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "82872221-aef0-5c8a-83c1-8c116c94a80b"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 7, "uuid": "82872221-aef0-5c8a-83c1-8c116c94a80b", "validation": {"axes": {"B": 1, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "82872221-aef0-5c8a-83c1-8c116c94a80b"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/cb5f723948a9906da254f5dbb706a9f2d9000dc1bcabae3d6af421242a63aab3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.0019972650334239006, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c48b85f2-bff4-5f3c-bfc7-e78c995ee073"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 7, "uuid": "c48b85f2-bff4-5f3c-bfc7-e78c995ee073", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00048828125], "status": "passed", "uuid": "c48b85f2-bff4-5f3c-bfc7-e78c995ee073"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/9aee725f4c705ab49598a8e1a119fe28ef20d39db9dd86db1a9575de9ac5c567.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.02804623544216156, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f849f163-b4b3-5616-84cd-83f86c0d1187"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 7, "uuid": "f849f163-b4b3-5616-84cd-83f86c0d1187", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "f849f163-b4b3-5616-84cd-83f86c0d1187"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/3b9304c0dabbf07cd0c66290fb61a372920e0d31da3021affc2c1ca1e91e32fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.004096947144716978, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9a5a6acb-75bc-5ab6-97d2-91d7c78c2483"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 7, "uuid": "9a5a6acb-75bc-5ab6-97d2-91d7c78c2483", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "9a5a6acb-75bc-5ab6-97d2-91d7c78c2483"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/89edf9a73c3f89eed8873ee550cfd0791251e2a95f6e01541f65cb63a64071bb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.006455938797444105, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5ecdfdff-3ac2-56c0-86b1-fe3a456700ee"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 7, "uuid": "5ecdfdff-3ac2-56c0-86b1-fe3a456700ee", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "5ecdfdff-3ac2-56c0-86b1-fe3a456700ee"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/d277927f0658865620cc5e24d1371e47c6bc998d782afd8ad24747b39925ee5c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.000244140625], "seconds": 0.006490407045930624, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b62e8d3a-b28f-5b6b-a9bc-51be5ad74e5e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 7, "uuid": "b62e8d3a-b28f-5b6b-a9bc-51be5ad74e5e", "validation": {"axes": {"B": 1, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.000244140625], "status": "passed", "uuid": "b62e8d3a-b28f-5b6b-a9bc-51be5ad74e5e"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 9, "X": 24, "Y": 14}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/c9a1b59ab2280beae723912c9ff8170322285e1141951c2142668d6eb4790694.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "d30833f7-16da-581b-84f4-66007ed77ae3", "validation": {"axes": {"B": 1, "T": 9, "X": 24, "Y": 14}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.04134256672114134, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d30833f7-16da-581b-84f4-66007ed77ae3"}, "variant": "grid_9x14x24", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 9, "X": 48, "Y": 28}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/7c155350a8970f49e963183abf8c263293b42601cea5d8c0a135bf48af359500.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "572aaf07-1f53-5279-b4d3-b8cc9ae232c8", "validation": {"axes": {"B": 1, "T": 9, "X": 48, "Y": 28}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.6162232710048556, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "572aaf07-1f53-5279-b4d3-b8cc9ae232c8"}, "variant": "grid_9x28x48", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 17, "X": 48, "Y": 28}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/7b03e5f130afa3425eeaddcd5651b0697b399999fa2d6824ca7db4a5026f23d4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "3eaae3b6-58d1-5ad4-94ba-ecf658aa5687", "validation": {"axes": {"B": 1, "T": 17, "X": 48, "Y": 28}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.23960432969033718, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3eaae3b6-58d1-5ad4-94ba-ecf658aa5687"}, "variant": "grid_17x28x48", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 33, "X": 96, "Y": 56}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/f70a1a8e117981439f190a4c6e55f60399f521bb45bf5a1bae82deee891b4e6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "b573c4cd-8e41-5e92-af6a-362e7b16f0a1", "validation": {"axes": {"B": 1, "T": 33, "X": 96, "Y": 56}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 1.8068100861273706, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b573c4cd-8e41-5e92-af6a-362e7b16f0a1"}, "variant": "grid_33x56x96", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 49, "X": 192, "Y": 112}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/6d86b779c94225e2ed9381b191d2ff2c71961c3afbb4ba80870d65d4f7bb8d02.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241711.9225547, "step": 0, "uuid": "bb8a6ae2-035a-51c1-b6a2-2aff689b88f1", "validation": {"axes": {"B": 1, "T": 49, "X": 192, "Y": 112}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 10.98438697680831, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bb8a6ae2-035a-51c1-b6a2-2aff689b88f1"}, "variant": "grid_49x112x192", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/6307cb0a22b7ac2e920fe9ad9453d89ef42f0b815f7cc695fe30c48ec9ad0395.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04800855414941907, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b76a8861-ae59-5f3b-92b4-c02019ca4941"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "b76a8861-ae59-5f3b-92b4-c02019ca4941", "validation": {"axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "b76a8861-ae59-5f3b-92b4-c02019ca4941"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/3f45356cd73e6d1045cf350d684ebac058e60d9b764ff34e85d9a092c6cf32f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0001220703125], "seconds": 0.001971811056137085, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f78b3e90-16eb-548e-b877-6d9f4d4d7acd"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "f78b3e90-16eb-548e-b877-6d9f4d4d7acd", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0001220703125], "status": "passed", "uuid": "f78b3e90-16eb-548e-b877-6d9f4d4d7acd"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/ff29237e311cf49dcd7b4bb16aefa60d36f0a0a66155b1519bb66c0ce8edecf4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.01644506072625518, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "43bb2e12-3ee6-5e23-9192-040a8405a7ae"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "43bb2e12-3ee6-5e23-9192-040a8405a7ae", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "43bb2e12-3ee6-5e23-9192-040a8405a7ae"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/87a4d3fd3fb1284882109437cd44ae9ecad1dc6ea5a4d2a88923aae98e2fbc87.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.000244140625], "seconds": 0.004052453674376011, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c59e00ad-a2d0-59a9-9ad8-8760960df7e2"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "c59e00ad-a2d0-59a9-9ad8-8760960df7e2", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.000244140625], "status": "passed", "uuid": "c59e00ad-a2d0-59a9-9ad8-8760960df7e2"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/a31fb2550e32724de7e7d65dd1c7e867acfe3122a2f4dc882cfa723bac70fdc5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.01087050512433052, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4522fae5-661b-5625-bad8-988b5cb005b1"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "4522fae5-661b-5625-bad8-988b5cb005b1", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "4522fae5-661b-5625-bad8-988b5cb005b1"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f4dafd2827123b850357c26784e25b2907e73366f2560227a238aa0884d620f7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.004464325029402971, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cacf8279-51df-5a97-a592-9d328a81df13"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "cacf8279-51df-5a97-a592-9d328a81df13", "validation": {"axes": {"B": 1, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "cacf8279-51df-5a97-a592-9d328a81df13"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/0b72eea4024de43033960bf52873fe4af465d06ad4432a952203952ef3675a77.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.017077457159757614, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "68caa03d-dfc2-5247-bee5-37a69aa2a52d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 4, "uuid": "68caa03d-dfc2-5247-bee5-37a69aa2a52d", "validation": {"axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "68caa03d-dfc2-5247-bee5-37a69aa2a52d"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/97783ed9bf6d72e13731c39ad639c4cb1d45cf2c2f6112157a0b53ac12fd8bfa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.001979278866201639, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aa0f37dd-bbb0-5ef5-9907-001d131f553c"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 4, "uuid": "aa0f37dd-bbb0-5ef5-9907-001d131f553c", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00048828125], "status": "passed", "uuid": "aa0f37dd-bbb0-5ef5-9907-001d131f553c"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/6326bc6250bb1e63da0b5a7e41bf586d09eae6c0da106d3b45b2e325cdace6ea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.016225908882915974, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "10493d50-a0f8-527f-8d39-414c377360f2"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 4, "uuid": "10493d50-a0f8-527f-8d39-414c377360f2", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "10493d50-a0f8-527f-8d39-414c377360f2"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/bc0371619e2c5f0ef19a3b71a8e1dd94c1ad8e5390824f7ca00d4148479eeb87.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.004029539879411459, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2a358cbf-1566-510c-b15e-274202339d18"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 4, "uuid": "2a358cbf-1566-510c-b15e-274202339d18", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00048828125], "status": "passed", "uuid": "2a358cbf-1566-510c-b15e-274202339d18"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/2125441614e33de8e5ad2326dca09b45cbeee76b779bd7a1160f4d3a6edd66fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.004312260076403618, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f300d2d2-eb25-5d81-b06b-2d25fbadd9ec"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 4, "uuid": "f300d2d2-eb25-5d81-b06b-2d25fbadd9ec", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "f300d2d2-eb25-5d81-b06b-2d25fbadd9ec"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/b571a1465fbdad5437b7a98dd5a36fe5ccab95b190660282338669ed7c16b3c4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.000244140625], "seconds": 0.004294415004551411, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8a1d591d-ebb7-57f3-8fd6-23897edf00a5"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 4, "uuid": "8a1d591d-ebb7-57f3-8fd6-23897edf00a5", "validation": {"axes": {"B": 1, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.000244140625], "status": "passed", "uuid": "8a1d591d-ebb7-57f3-8fd6-23897edf00a5"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/429b37ccea1cf717aa82f3e5c5799423ba77305a514a6852fbff7f34f18f65b5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.014925865922123194, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "328c3d7b-76a2-5bf9-a9be-cedd6c2c58d8"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 7, "uuid": "328c3d7b-76a2-5bf9-a9be-cedd6c2c58d8", "validation": {"axes": {"B": 1, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "328c3d7b-76a2-5bf9-a9be-cedd6c2c58d8"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/c9fcc64470f06cff7542376da19817b9652a7f1a8f5840bfdc42cc19c023efe7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0019850628450512886, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8867a364-4a86-5118-aeeb-e8adc8419445"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 7, "uuid": "8867a364-4a86-5118-aeeb-e8adc8419445", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "8867a364-4a86-5118-aeeb-e8adc8419445"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/d83108a7a8aabfb7f213b2bef649506910058d15992121a25fd8aea46c827a2b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.013812421821057796, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ca1ac853-8f3e-5179-902b-06ef7d561703"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 7, "uuid": "ca1ac853-8f3e-5179-902b-06ef7d561703", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "ca1ac853-8f3e-5179-902b-06ef7d561703"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/a9dae4c8c01ead85bf4cb6c39b36646cde3b749278991a27c56b19fd75f320b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [6.103515625e-05], "seconds": 0.004102291073650122, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8471e8e5-d50c-51c2-b946-b92561322441"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 7, "uuid": "8471e8e5-d50c-51c2-b946-b92561322441", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [6.103515625e-05], "status": "passed", "uuid": "8471e8e5-d50c-51c2-b946-b92561322441"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/65fd3ad1671d5e277942a3ed73691dba3640e18a4c47f2123563a83f6f5d4883.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004176128190010786, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cd816f58-b2a2-5a9f-bd83-410109d2df31"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 7, "uuid": "cd816f58-b2a2-5a9f-bd83-410109d2df31", "validation": {"axes": {"B": 1, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "cd816f58-b2a2-5a9f-bd83-410109d2df31"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/45a46e9b060a9c2765a49b17cd42d8b88a625352c31cb9c9b7486e6130197e48.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [3.0517578125e-05], "seconds": 0.0041802311316132545, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e2529ef9-37e8-5fe2-98ef-82136023e5a9"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 7, "uuid": "e2529ef9-37e8-5fe2-98ef-82136023e5a9", "validation": {"axes": {"B": 1, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [3.0517578125e-05], "status": "passed", "uuid": "e2529ef9-37e8-5fe2-98ef-82136023e5a9"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 9, "X": 16, "Y": 10}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/75547a2c62a7b00a958343f9e6d171d057326dd4f776bf04664b8147240913ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "503c803e-a524-562c-a980-daf2bad42368", "validation": {"axes": {"B": 1, "T": 9, "X": 16, "Y": 10}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.02088306611403823, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "503c803e-a524-562c-a980-daf2bad42368"}, "variant": "grid_9x10x16", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 9, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/e405f2f1bb5a266786f48ac34162d6cbbee87032aff59edcd854d4fb16b7539a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "ba0a26ae-0688-5564-a415-b40944dc1d2c", "validation": {"axes": {"B": 1, "T": 9, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.06787523115053773, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ba0a26ae-0688-5564-a415-b40944dc1d2c"}, "variant": "grid_9x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 17, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/abae80365021b2345eba899010dd61cba48c1e1f7fc2572a9f484c44d76a6002.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "95130464-5fd0-59c3-81b3-5759675a89f0", "validation": {"axes": {"B": 1, "T": 17, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.11517253704369068, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "95130464-5fd0-59c3-81b3-5759675a89f0"}, "variant": "grid_17x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 33, "X": 64, "Y": 40}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/a9ed6a3e5469db99b6fe4f1a95c9e8bb6de1784ba4da869ec3be1a54ed57fb36.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "d25c33ca-dcd8-58c5-b326-9a5e77c4d857", "validation": {"axes": {"B": 1, "T": 33, "X": 64, "Y": 40}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.8816024870611727, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d25c33ca-dcd8-58c5-b326-9a5e77c4d857"}, "variant": "grid_33x40x64", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 49, "X": 128, "Y": 80}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/9963ed274554cd6e176d498a07bf093818db8e4f015d2a5572cec7d92c63a8db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241758.0054827, "step": 0, "uuid": "4bc37940-bacd-56bd-b86c-51279edc24de", "validation": {"axes": {"B": 1, "T": 49, "X": 128, "Y": 80}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 4.989393078256398, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4bc37940-bacd-56bd-b86c-51279edc24de"}, "variant": "grid_49x80x128", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/e0841729cb92b1cd9d247c730153dca1fdb40c57bfd14950d2df9db6bac60946.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.054834083653986454, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "22d5bea1-549f-5d31-bb56-4be38b775291"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "22d5bea1-549f-5d31-bb56-4be38b775291", "validation": {"axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "22d5bea1-549f-5d31-bb56-4be38b775291"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/283148eb29dff8a7dc71d586457543be9f86bcfc0480839c0414d3a52f25c263.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.0029808501712977886, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6c5c668c-0b54-5593-b57a-cf2a4c681609"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "6c5c668c-0b54-5593-b57a-cf2a4c681609", "validation": {"axes": {"B": 1, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "6c5c668c-0b54-5593-b57a-cf2a4c681609"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/a5a4b72812772ea31d92209ed871796330b88a27474854fc7ba095b6ab224e72.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.031006361823529005, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "93a38d68-1ec7-54cf-9bf8-0d46651c3adf"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "93a38d68-1ec7-54cf-9bf8-0d46651c3adf", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "93a38d68-1ec7-54cf-9bf8-0d46651c3adf"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/2e6649a91720ca8a911e07694adb93856a6499c0f36cd81d07ddda0ebecb39e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.004909697920084, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "69164a63-b5fb-5259-9dde-e99bac5d08c3"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "69164a63-b5fb-5259-9dde-e99bac5d08c3", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "69164a63-b5fb-5259-9dde-e99bac5d08c3"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/bfd39246a5c540352adc8afa6f1f209025ada95bfa033c80a6217f07e8ee61e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.009198926389217377, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "67d71bcc-ecf0-57a2-8af2-2de9f591c6e0"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "67d71bcc-ecf0-57a2-8af2-2de9f591c6e0", "validation": {"axes": {"B": 1, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "67d71bcc-ecf0-57a2-8af2-2de9f591c6e0"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/358a8705e9de07aaa9dcd376864dd3dc7195f03bb6bb85d20bd05a171274b58e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.008517648093402386, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "53d14356-3ab0-531c-9e45-2d0d3f400f2e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "53d14356-3ab0-531c-9e45-2d0d3f400f2e", "validation": {"axes": {"B": 1, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "53d14356-3ab0-531c-9e45-2d0d3f400f2e"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/4ed745038d995b1f7b638e5634151b8e63e940bb60d5cb2c9623d8dd122ddd09.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.04977079015225172, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a34594b2-2c8d-58cd-b443-630151fec624"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 4, "uuid": "a34594b2-2c8d-58cd-b443-630151fec624", "validation": {"axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "a34594b2-2c8d-58cd-b443-630151fec624"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/42e18ada4d1fe2c844bb52be33a591cdc7ecadf813a2082a7efe7a8da505b818.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0024036471731960773, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e4210683-d2f7-5342-b764-9ce466e995a7"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 4, "uuid": "e4210683-d2f7-5342-b764-9ce466e995a7", "validation": {"axes": {"B": 1, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "e4210683-d2f7-5342-b764-9ce466e995a7"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/a02b340cc222492791888e9cff78c52833331793121a2c78dcfcbdc6e6b999eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.030632327776402235, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a5c07816-ebf1-5d84-b114-768db1f7f14a"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 4, "uuid": "a5c07816-ebf1-5d84-b114-768db1f7f14a", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "a5c07816-ebf1-5d84-b114-768db1f7f14a"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/fe742b385cbf48c01ae1753de75ce290d897929df8927ede07b8bcd635b7ca05.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.0048117199912667274, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "93257e9c-2c1a-5b9f-8701-3efd87e45bdb"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 4, "uuid": "93257e9c-2c1a-5b9f-8701-3efd87e45bdb", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00048828125], "status": "passed", "uuid": "93257e9c-2c1a-5b9f-8701-3efd87e45bdb"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/c56548ccd797ef9eb59f6176c00b2a85175896db33c41a02b347ea639510d460.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.008404413238167763, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b89bf097-d8ff-546b-bdf0-c61be5f0a31c"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 4, "uuid": "b89bf097-d8ff-546b-bdf0-c61be5f0a31c", "validation": {"axes": {"B": 1, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "b89bf097-d8ff-546b-bdf0-c61be5f0a31c"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/d57b6686c58237d96bb4a6848142e81cc33ece56196775a5ddf697b206628b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.008421491831541061, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8e10d048-f490-55b5-8b20-6b27d5367a30"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 4, "uuid": "8e10d048-f490-55b5-8b20-6b27d5367a30", "validation": {"axes": {"B": 1, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "8e10d048-f490-55b5-8b20-6b27d5367a30"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/493a5a17bedd061b42b22f951d095f0e2e1e72896e3e22e279b19d59b90c3950.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.04946383694186807, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "325e1a4d-dd1b-5634-af96-f4344393a4b3"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 7, "uuid": "325e1a4d-dd1b-5634-af96-f4344393a4b3", "validation": {"axes": {"B": 1, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "325e1a4d-dd1b-5634-af96-f4344393a4b3"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f4ec26669e5826b595ccdc08cf9ca1e6038e9acf2f65e51052e6551e760f68c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.0023901108652353287, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bc2db615-2925-5e46-8a5d-9d6c3a192d50"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 7, "uuid": "bc2db615-2925-5e46-8a5d-9d6c3a192d50", "validation": {"axes": {"B": 1, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "bc2db615-2925-5e46-8a5d-9d6c3a192d50"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/deabdb22bb17649c96f781e5cdfd8e7e4ca40d8d7c57de76cd2e0666b9b436b7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.030247369315475225, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "27af7d20-3b72-5881-ab32-4a3ab8d2cdd0"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 7, "uuid": "27af7d20-3b72-5881-ab32-4a3ab8d2cdd0", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "27af7d20-3b72-5881-ab32-4a3ab8d2cdd0"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/01d4ebb18d59fd4e558fcbc47e2996ddd8550e9a92b9313e199ef8f19eb5876e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.004569757264107466, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ffc1ed3e-8a2f-5ad4-8413-4ac86f2422f1"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 7, "uuid": "ffc1ed3e-8a2f-5ad4-8413-4ac86f2422f1", "validation": {"axes": {"B": 1, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00048828125], "status": "passed", "uuid": "ffc1ed3e-8a2f-5ad4-8413-4ac86f2422f1"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/5e6ebfd69531b666fd55597233632b55bdd03c58c8da6c80af18d5ec7a9c548e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.007825418375432491, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "492f0684-6444-57cf-b9ab-f56d851b6f13"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 7, "uuid": "492f0684-6444-57cf-b9ab-f56d851b6f13", "validation": {"axes": {"B": 1, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "492f0684-6444-57cf-b9ab-f56d851b6f13"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/47db632032bdbf9839f003e9db86067f1342f3ddc7f8293f19edb48c15129589.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 1, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.007775096222758293, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "eae3dd17-7680-5eda-8915-1a36ed2f03b0"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 7, "uuid": "eae3dd17-7680-5eda-8915-1a36ed2f03b0", "validation": {"axes": {"B": 1, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "eae3dd17-7680-5eda-8915-1a36ed2f03b0"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 18, "X": 16, "Y": 10}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/6c4021dc447efc2f8b874bf19d7c29b7c6cb9c84bf5a91c34183ed042d9a6cd2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "d9ce70f4-0314-530a-a4d3-cb359744d16b", "validation": {"axes": {"B": 1, "T": 18, "X": 16, "Y": 10}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.039467715192586184, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d9ce70f4-0314-530a-a4d3-cb359744d16b"}, "variant": "grid_18x10x16", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 18, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/b6aa794d0622e0d7c9d7ee8dcb39c463ef97770d97e1595d55dfb1357b50e326.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "5b35db05-d838-5b18-9415-11dcf35d9cff", "validation": {"axes": {"B": 1, "T": 18, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.13199147302657366, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5b35db05-d838-5b18-9415-11dcf35d9cff"}, "variant": "grid_18x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 35, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/cdb619e744bcec546fff6de2bac95d97ce850b0aff47ddc3a9ae580ff8695dc9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "65bd58a7-ae16-5b36-82e1-7c801c494fc3", "validation": {"axes": {"B": 1, "T": 35, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.23392906924709678, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "65bd58a7-ae16-5b36-82e1-7c801c494fc3"}, "variant": "grid_35x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 69, "X": 64, "Y": 40}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/0546fe692e1c0d4bff54383bca8f443011a4281afd7727698e5ced6e0def8653.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "fdcc196f-3464-5db0-b64f-f2e29fe569fa", "validation": {"axes": {"B": 1, "T": 69, "X": 64, "Y": 40}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 1.7970105661079288, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fdcc196f-3464-5db0-b64f-f2e29fe569fa"}, "variant": "grid_69x40x64", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 1, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 1, "T": 121, "X": 128, "Y": 80}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/3af97dce0be5efa434edbee42cfb9793ce2e9d56410836623eafaa5f6355ed60.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 1, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video", "requested_outputs": 1, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b1-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b1-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241772.0388129, "step": 0, "uuid": "f0fe22fd-bde6-5d95-b486-a1233f5342de", "validation": {"axes": {"B": 1, "T": 121, "X": 128, "Y": 80}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 12.406822388060391, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f0fe22fd-bde6-5d95-b486-a1233f5342de"}, "variant": "grid_121x80x128", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/ba010bc4d143583f885320420cda12789bc2e572d95b0a28906c7222ae587e65.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.3564747399650514, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6c3637a4-7d4e-515c-9f0b-eeaf53ed75ca"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "6c3637a4-7d4e-515c-9f0b-eeaf53ed75ca", "validation": {"axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "6c3637a4-7d4e-515c-9f0b-eeaf53ed75ca"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/3010017dac0ec793a8e7587561d111ab067160e2d1ef921edaa5424f582b7e16.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0001220703125], "seconds": 4.4906132887117565, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8e822ec0-b226-5b54-a3fa-ffce82467b80"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "8e822ec0-b226-5b54-a3fa-ffce82467b80", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0001220703125], "status": "passed", "uuid": "8e822ec0-b226-5b54-a3fa-ffce82467b80"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/8aab9a9ed5a55d42f08e4d28d07e43177c927d5bd53db3b1b677154c6005af12.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.1707967766560614, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d617c58a-373a-5b40-9e36-e7c009786df5"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "d617c58a-373a-5b40-9e36-e7c009786df5", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "d617c58a-373a-5b40-9e36-e7c009786df5"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/aa7c481b31ca35f3ae086bed2d8f7a4369ae6499749e53b3aa42217a73065653.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.015585278160870075, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "72cf79c8-d237-5ede-b05b-111959bbd1b9"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "72cf79c8-d237-5ede-b05b-111959bbd1b9", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "72cf79c8-d237-5ede-b05b-111959bbd1b9"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/097602210ffcc0e9d882f3b7706633a2d97decb565d197c5a1132329f734bab9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.030875524040311575, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b00ce07f-9e07-5f83-a74b-54a97fe37b77"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "b00ce07f-9e07-5f83-a74b-54a97fe37b77", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "b00ce07f-9e07-5f83-a74b-54a97fe37b77"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/7573886bd67a7f544f14b46babc98f2339aa19f633be64c09dff0e9338f54fe6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.011705226730555296, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bf4a3d53-aa18-5df0-b5c8-8a058ab5ec36"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "bf4a3d53-aa18-5df0-b5c8-8a058ab5ec36", "validation": {"axes": {"B": 2, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "bf4a3d53-aa18-5df0-b5c8-8a058ab5ec36"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/f6c4bc7ef6f8cc8bbda7b0626b911f4e3ca2122928f67bb46312f8517c544b0f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.08746376121416688, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9f69a6cc-1028-52c5-8516-12d5f44e17ea"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 4, "uuid": "9f69a6cc-1028-52c5-8516-12d5f44e17ea", "validation": {"axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "9f69a6cc-1028-52c5-8516-12d5f44e17ea"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/12dc05d41935be6b0e5597c78ca18e9a1f365c99ba7c5a1f70c4033050c2d3c2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0023189890198409557, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "569ab1b0-9ec5-5303-ba3e-c3d7e6a4f769"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 4, "uuid": "569ab1b0-9ec5-5303-ba3e-c3d7e6a4f769", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "569ab1b0-9ec5-5303-ba3e-c3d7e6a4f769"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/6381269f3a80c566184b5cbee5460bdaf10e31bad9bf06cefe695978c98aeeee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.056300828233361244, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "06131807-accf-5812-a37b-8f9ae948e0f5"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 4, "uuid": "06131807-accf-5812-a37b-8f9ae948e0f5", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "06131807-accf-5812-a37b-8f9ae948e0f5"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/6d53eff0bac33b16776ea563afd8fcad726e3fa7b6f27b05c7a0b2cb968eff06.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0063310349360108376, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "caf929e4-a070-5ef3-aed8-5a916892b20f"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 4, "uuid": "caf929e4-a070-5ef3-aed8-5a916892b20f", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "caf929e4-a070-5ef3-aed8-5a916892b20f"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/febcaa41d7b8d26203636b176e2c36818718b6005269f6ba962646337a105497.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.01148474495857954, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c7c4277a-2524-59b9-98c8-3422125657eb"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 4, "uuid": "c7c4277a-2524-59b9-98c8-3422125657eb", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "c7c4277a-2524-59b9-98c8-3422125657eb"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/d7a7b40eb46f6a75bfb7add8dc3cf774115c4a98f1183361dedaea6fb5571fd0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.011359761003404856, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "df7ab88b-e699-519b-9c1c-ab1b37699c14"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 4, "uuid": "df7ab88b-e699-519b-9c1c-ab1b37699c14", "validation": {"axes": {"B": 2, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "df7ab88b-e699-519b-9c1c-ab1b37699c14"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/55d34f8ed230e27f4420fa9499b9f182a87c0dcbb359ebed2c572d24d59267cd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.08758193673565984, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1802abf8-29bb-5236-85dd-195a445b9ab2"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 7, "uuid": "1802abf8-29bb-5236-85dd-195a445b9ab2", "validation": {"axes": {"B": 2, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "1802abf8-29bb-5236-85dd-195a445b9ab2"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/7c401218dc23f058af594381bf2ce51db4a1ae46d2fc3be5588cdfec3c58e0eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0022568400017917156, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fe5368a5-d8b1-53b5-a6c7-24893c1a525a"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 7, "uuid": "fe5368a5-d8b1-53b5-a6c7-24893c1a525a", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "fe5368a5-d8b1-53b5-a6c7-24893c1a525a"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/a5f55d6f057b45ab8e6cfbf3349834a86beed72b38e27fc01db1c1ca28e205e2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.055854735895991325, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d7ed325e-48af-50dd-a16a-d93ebe4d1926"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 7, "uuid": "d7ed325e-48af-50dd-a16a-d93ebe4d1926", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "d7ed325e-48af-50dd-a16a-d93ebe4d1926"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/e35850b82a4fde3f2c9f2b44030dc409695d60887543fa082adb2dfa0c55075e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.006105829030275345, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c8a4b5bf-8202-51c8-9496-a3376915f766"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 7, "uuid": "c8a4b5bf-8202-51c8-9496-a3376915f766", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "c8a4b5bf-8202-51c8-9496-a3376915f766"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/57c7b2b1eb91ba22fd2a6efd669eb663c68d4a42a9effa03f6040e469ea9332f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010980389080941677, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9d40f2b5-a110-5aae-a02b-bf27acc4f3f0"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 7, "uuid": "9d40f2b5-a110-5aae-a02b-bf27acc4f3f0", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "9d40f2b5-a110-5aae-a02b-bf27acc4f3f0"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/4f0dacfd1e56a365d4c30cf0f6e9c92bf3eee0c11c12e8a3d185cb164cfacbb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.010989570058882236, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd58f2b7-9e92-58a5-851f-900227514d7c"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 7, "uuid": "bd58f2b7-9e92-58a5-851f-900227514d7c", "validation": {"axes": {"B": 2, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "bd58f2b7-9e92-58a5-851f-900227514d7c"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 9, "X": 24, "Y": 14}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/fe3fb7acba9665f059ceaf13ed5e823d9c90686316279df7a6bf07f795039cd5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "01fc8439-9178-5ae6-b4a9-8a81d737d807", "validation": {"axes": {"B": 2, "T": 9, "X": 24, "Y": 14}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.06161692691966891, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "01fc8439-9178-5ae6-b4a9-8a81d737d807"}, "variant": "grid_9x14x24", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 9, "X": 48, "Y": 28}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/98c5b9560f3d9a8ba3133adf125a9d11be01a47b739660f4940c27987d7b2cb1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "d6c3cd98-bcbb-5014-96a6-e228b4157aab", "validation": {"axes": {"B": 2, "T": 9, "X": 48, "Y": 28}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.6253014388494194, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d6c3cd98-bcbb-5014-96a6-e228b4157aab"}, "variant": "grid_9x28x48", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 17, "X": 48, "Y": 28}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/8d0b7c04842036e5b642f5c9e746897d9de74bb3eda689858beb4e3f908e1dfb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "7a5a1b75-2d8f-587c-842a-fdd539e9f60e", "validation": {"axes": {"B": 2, "T": 17, "X": 48, "Y": 28}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.2609353680163622, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7a5a1b75-2d8f-587c-842a-fdd539e9f60e"}, "variant": "grid_17x28x48", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 33, "X": 96, "Y": 56}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/3b7a1d71546e723e1c9a9969f524109f46005fc85d6b667d6e0e83d00e0630e2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "09b0007c-487c-5337-91cb-b6b9f153a218", "validation": {"axes": {"B": 2, "T": 33, "X": 96, "Y": 56}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 1.9233809900470078, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "09b0007c-487c-5337-91cb-b6b9f153a218"}, "variant": "grid_33x56x96", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 49, "X": 192, "Y": 112}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/b0b59766fd7c1781275bb7e5029da80ad4bfaf51b9c21e8452e19b36bd9b9d0d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792241953.9250946, "step": 0, "uuid": "7df22f0a-44c5-5824-a3c7-8789320c1fcc", "validation": {"axes": {"B": 2, "T": 49, "X": 192, "Y": 112}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 18.63961925683543, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7df22f0a-44c5-5824-a3c7-8789320c1fcc"}, "variant": "grid_49x112x192", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/adec87acc9de1d195c2646cb04593537ef8a01365768d2394eafe16dd999e991.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.09061764599755406, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "695ccefb-838b-52c6-8d57-7e5f6a5b755f"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "695ccefb-838b-52c6-8d57-7e5f6a5b755f", "validation": {"axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "695ccefb-838b-52c6-8d57-7e5f6a5b755f"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/3010017dac0ec793a8e7587561d111ab067160e2d1ef921edaa5424f582b7e16.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0001220703125], "seconds": 0.0022692279890179634, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cae9e5c2-dc82-5497-90bc-0fb61d58d62c"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "cae9e5c2-dc82-5497-90bc-0fb61d58d62c", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0001220703125], "status": "passed", "uuid": "cae9e5c2-dc82-5497-90bc-0fb61d58d62c"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/52b2132db1569942dd9adedca02a096ee6100ba701a3c1bc82fbc0fc0b9d4414.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03103246120736003, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "72c5ddc0-a738-575e-9cce-35d0c4d7430a"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "72c5ddc0-a738-575e-9cce-35d0c4d7430a", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "72c5ddc0-a738-575e-9cce-35d0c4d7430a"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/5e7d4c5c27a95acc9a33ea4d8532a641856b2325565a9caaf3e3a9b3f6f189cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.006090428214520216, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9c5eb397-8a6c-53a6-87b5-e36789679d66"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "9c5eb397-8a6c-53a6-87b5-e36789679d66", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "9c5eb397-8a6c-53a6-87b5-e36789679d66"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/bd87b065b49c346452a46e3b989ccad0251500682c241dff27355ff17be9ffc8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.014651388861238956, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bf743591-8506-5869-9e16-be6e8cb3aa4e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "bf743591-8506-5869-9e16-be6e8cb3aa4e", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "bf743591-8506-5869-9e16-be6e8cb3aa4e"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/4a340cfd7e2e349ade624205639aec26f9f10f9a2ddb404e55fe0c72cfbb7bc7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.009273265022784472, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9b9f5473-8f78-599e-af55-b83dff6649b0"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "9b9f5473-8f78-599e-af55-b83dff6649b0", "validation": {"axes": {"B": 2, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "9b9f5473-8f78-599e-af55-b83dff6649b0"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/d0466745315073f3b900499277298e8c4df7d035a7291d91365a8b7d57681f60.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03225407283753157, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6e7f3c9e-ea8c-519d-b268-8c1b5c19a937"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 4, "uuid": "6e7f3c9e-ea8c-519d-b268-8c1b5c19a937", "validation": {"axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "6e7f3c9e-ea8c-519d-b268-8c1b5c19a937"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/4f09df34bbb4148be0197189046408ea112b188af42fad11845cb4a9354b6c51.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0022834837436676025, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bc426141-767c-545e-94e2-4ab7ef1ad6ea"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 4, "uuid": "bc426141-767c-545e-94e2-4ab7ef1ad6ea", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "bc426141-767c-545e-94e2-4ab7ef1ad6ea"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/624f3b4068ccadfc3e81d536c052bd10aa3d4eedf2184474f425bdcc92225f9c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.030449194833636284, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b7b2749e-15a7-5e24-a748-28c0329e6220"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 4, "uuid": "b7b2749e-15a7-5e24-a748-28c0329e6220", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "b7b2749e-15a7-5e24-a748-28c0329e6220"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/03b3dbdfe994e8e6caeabc055ea6872581cc5ef594c7c7bd3a81291d7929d70b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.006085486151278019, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f399b280-d1ec-599a-9a39-69f063be5ee8"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 4, "uuid": "f399b280-d1ec-599a-9a39-69f063be5ee8", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "f399b280-d1ec-599a-9a39-69f063be5ee8"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/41b97cd6a0b571e179b13780c968a80425800222341f2b4cfd7c0278b3ccc05c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.006569539196789265, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "475965cf-cbad-5baa-b685-479860697719"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 4, "uuid": "475965cf-cbad-5baa-b685-479860697719", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "475965cf-cbad-5baa-b685-479860697719"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/774bbbf14aa4bfd894e5c2f09f3a4fb555bdecfbf57bc7ec8c152a9749f114f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [6.103515625e-05], "seconds": 0.00666813226416707, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2f06424d-4917-55f1-a49e-92062ee6393d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 4, "uuid": "2f06424d-4917-55f1-a49e-92062ee6393d", "validation": {"axes": {"B": 2, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [6.103515625e-05], "status": "passed", "uuid": "2f06424d-4917-55f1-a49e-92062ee6393d"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/711bed213b009a600243d4f530e2d10a817da768e604f5b0204f1866c5f37f1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.030416897032409906, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "79c92abe-237a-524b-8619-6a2084a92d6d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 7, "uuid": "79c92abe-237a-524b-8619-6a2084a92d6d", "validation": {"axes": {"B": 2, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "79c92abe-237a-524b-8619-6a2084a92d6d"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/5018049bb7e1a47e6d05a84befd96df1bea51b8357a3d2ba52ef9c9b41c6b186.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.0022522597573697567, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "56cd6408-c046-526d-81a1-7e25013dfe00"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 7, "uuid": "56cd6408-c046-526d-81a1-7e25013dfe00", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "56cd6408-c046-526d-81a1-7e25013dfe00"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/6282ff2b4e962b8ed0b5efd725fb77a70afc9c2f04daf7fc7ddf59359da05c56.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.028038349002599716, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d48a99f0-024c-518f-9eb2-b0fdf84a524d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 7, "uuid": "d48a99f0-024c-518f-9eb2-b0fdf84a524d", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "d48a99f0-024c-518f-9eb2-b0fdf84a524d"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/e413095fe96f6bd2e4d23a59c75999ff031b7c391c7f1f3ce6afd72ddcc09a2a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.006088613998144865, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3e7b0302-b42e-5674-8b89-9c9bc8cef185"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 7, "uuid": "3e7b0302-b42e-5674-8b89-9c9bc8cef185", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "3e7b0302-b42e-5674-8b89-9c9bc8cef185"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/c0c4f8305404571a75783a3bcd1befb00f7edaefdf84296ab6f10d0386e3d65a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.006354016251862049, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0c4ac2c3-4269-569a-a7d8-24a40c88d6c8"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 7, "uuid": "0c4ac2c3-4269-569a-a7d8-24a40c88d6c8", "validation": {"axes": {"B": 2, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "0c4ac2c3-4269-569a-a7d8-24a40c88d6c8"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/9abb0dad74d23a188f93f3c017513dc4ab050c0dbebf5a01637ba5fb61601556.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.006424283143132925, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e251f5a0-c781-58a5-951e-21de1acf7668"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 7, "uuid": "e251f5a0-c781-58a5-951e-21de1acf7668", "validation": {"axes": {"B": 2, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "e251f5a0-c781-58a5-951e-21de1acf7668"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 9, "X": 16, "Y": 10}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/a82e4ffb152558a9bf267419597ab378797101958d6bf7530ba1f25b6686a2b9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "3b569bfb-00e7-594e-a79b-87cb68e3a016", "validation": {"axes": {"B": 2, "T": 9, "X": 16, "Y": 10}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.030310614965856075, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3b569bfb-00e7-594e-a79b-87cb68e3a016"}, "variant": "grid_9x10x16", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 9, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/86e79d29e9f09952cf348b7ed7a767e7be06c9b63d12bae34a1f6d5b510ac196.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "06524fdc-1583-57f2-80e2-fdbbe82c11cf", "validation": {"axes": {"B": 2, "T": 9, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.07690511969849467, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "06524fdc-1583-57f2-80e2-fdbbe82c11cf"}, "variant": "grid_9x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 17, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/97e2496f90361ab146e74eaf5884427a79aedd5b705cbb34a5361d3baf42aeb2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "1a60d4e5-bf8d-509a-aa67-96eb03bd8e7e", "validation": {"axes": {"B": 2, "T": 17, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.12507362896576524, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a60d4e5-bf8d-509a-aa67-96eb03bd8e7e"}, "variant": "grid_17x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 33, "X": 64, "Y": 40}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/bd4774e3d42e8794030001de6ba2c474d0ea76bac76a973e3ecb36fb2a37a3e2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "6ed37cda-b6d1-57c6-bbd2-5ef073b679e1", "validation": {"axes": {"B": 2, "T": 33, "X": 64, "Y": 40}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.9556439155712724, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ed37cda-b6d1-57c6-bbd2-5ef073b679e1"}, "variant": "grid_33x40x64", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 49, "X": 128, "Y": 80}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/fc0616dacda9dca25e9fdaba0d6bbf49bc9849fef05d748db5e65bf793fb2d98.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242006.0334532, "step": 0, "uuid": "3aacba35-3f6d-5c90-8c63-a78396ea8d44", "validation": {"axes": {"B": 2, "T": 49, "X": 128, "Y": 80}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 8.43686280399561, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3aacba35-3f6d-5c90-8c63-a78396ea8d44"}, "variant": "grid_49x80x128", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/fd7c3482e9774d41d70a79cebfbb253b823b857809a39f22bdb6101db8f97aa3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.10575728584080935, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "195ec930-50e5-5ab2-b6af-ce9bcede4f65"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "195ec930-50e5-5ab2-b6af-ce9bcede4f65", "validation": {"axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "195ec930-50e5-5ab2-b6af-ce9bcede4f65"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/c446eefba96bda4d06adba9c9347bf0ae90d330135033d1cdbb1c02ca03b9c5e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.0038154777139425278, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1f9ae2ca-2f03-59fb-aa8a-bdaa967c3513"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "1f9ae2ca-2f03-59fb-aa8a-bdaa967c3513", "validation": {"axes": {"B": 2, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "1f9ae2ca-2f03-59fb-aa8a-bdaa967c3513"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/dbf89936000795742b734c37465068c640828da8e8f8b23fa28e429dd07e6eba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.06242698756977916, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "14f218bb-57ad-5372-a8fe-18fa938c250a"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "14f218bb-57ad-5372-a8fe-18fa938c250a", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "14f218bb-57ad-5372-a8fe-18fa938c250a"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f460603b964888235d48514813718de8ecb0b9e02b9849767033cf9c9b1be285.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.00763657595962286, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "12c50ef4-b8c1-5f90-88e3-8721a5aa5c70"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "12c50ef4-b8c1-5f90-88e3-8721a5aa5c70", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "12c50ef4-b8c1-5f90-88e3-8721a5aa5c70"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/393df4e7dc3bf41de7a9a06cc096c3eee5592816e29a38d8e1e66930d8f15eea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.01616377802565694, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ac82acc5-5fed-5aad-aef3-f63bf7050267"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "ac82acc5-5fed-5aad-aef3-f63bf7050267", "validation": {"axes": {"B": 2, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "ac82acc5-5fed-5aad-aef3-f63bf7050267"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/0cd13e0ddadf9f0510a5cd7115e948fa521a1858e817840479221aa45f0e1f8d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.01491745188832283, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "35d00eff-429f-5302-9ebb-a3d1971a22cc"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "35d00eff-429f-5302-9ebb-a3d1971a22cc", "validation": {"axes": {"B": 2, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "35d00eff-429f-5302-9ebb-a3d1971a22cc"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/65a0cdfa2530d0f36fb27016331d1cdb1615d6d3fa2d70a86bd8201d2682f2ad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.10005606804043055, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5188525c-ae23-5e83-aaa0-7233b2d1be4d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 4, "uuid": "5188525c-ae23-5e83-aaa0-7233b2d1be4d", "validation": {"axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "5188525c-ae23-5e83-aaa0-7233b2d1be4d"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/cc41afb741391a57f878fddf7a0699bb25bc923e8a3de79fca126859c5205583.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003065758850425482, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b33ed0a6-33a1-5651-ae01-36b17523f6aa"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 4, "uuid": "b33ed0a6-33a1-5651-ae01-36b17523f6aa", "validation": {"axes": {"B": 2, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "b33ed0a6-33a1-5651-ae01-36b17523f6aa"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/0b2803a8cb4a1390c3e3ad4a503413c32981793a1a7a7f000fb3687e8d5bb4e3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.06197610730305314, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b63f6ceb-944d-506e-97a6-aef2e8e454d8"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 4, "uuid": "b63f6ceb-944d-506e-97a6-aef2e8e454d8", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "b63f6ceb-944d-506e-97a6-aef2e8e454d8"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/c3f25e175f8c5dfb3f7eb2c4fd833aa02233c9073cd5a76160bbfa9a203d5e7e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.007583353202790022, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "acabbb4d-4ad9-59bd-ad85-aa59e5c61616"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 4, "uuid": "acabbb4d-4ad9-59bd-ad85-aa59e5c61616", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "acabbb4d-4ad9-59bd-ad85-aa59e5c61616"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/1a89097b8367073fbb8343344b969f961fce6f7a58bd99da28935f3fb1ebde72.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.014985054731369019, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "eaf051f7-e6f3-5dfe-8883-293fdc53d10e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 4, "uuid": "eaf051f7-e6f3-5dfe-8883-293fdc53d10e", "validation": {"axes": {"B": 2, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "eaf051f7-e6f3-5dfe-8883-293fdc53d10e"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/1926d870bb9ed06fa070059ee3d66824f667ca72baf513b715ff0350248fbf12.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.014807262923568487, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0aee85e5-f45c-5574-ba1f-84fa3e8bfb6d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 4, "uuid": "0aee85e5-f45c-5574-ba1f-84fa3e8bfb6d", "validation": {"axes": {"B": 2, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "0aee85e5-f45c-5574-ba1f-84fa3e8bfb6d"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/5e1577f1b2aee0f9321d34135db5adec510e70c4ff08787b6d6aca8827a20800.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0625], "seconds": 0.0994175998494029, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e52e3dab-ac58-5101-ab27-5bdae9bc911d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 7, "uuid": "e52e3dab-ac58-5101-ab27-5bdae9bc911d", "validation": {"axes": {"B": 2, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0625], "status": "passed", "uuid": "e52e3dab-ac58-5101-ab27-5bdae9bc911d"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/7847fa735a8f8c5156dc88dc9526883dd7d2320cfe8679076557f4c30b4359e7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.00308423675596714, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2c6d44e7-ea42-5814-8813-6a8164649bc4"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 7, "uuid": "2c6d44e7-ea42-5814-8813-6a8164649bc4", "validation": {"axes": {"B": 2, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "2c6d44e7-ea42-5814-8813-6a8164649bc4"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/9f3460de3afd3264471b55ad9218e76a58e1bd79366d9bcf89689616f80e3293.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06198076577857137, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cbe34697-5e71-573c-a820-8ff5d93ccf3b"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 7, "uuid": "cbe34697-5e71-573c-a820-8ff5d93ccf3b", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "cbe34697-5e71-573c-a820-8ff5d93ccf3b"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/d649fa3970e8ddac8022e57ade6c44336a37ac555d1981a0c16ca0aee443ea11.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.007042011711746454, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4795bcdd-6f37-5076-9404-a44be707081c"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 7, "uuid": "4795bcdd-6f37-5076-9404-a44be707081c", "validation": {"axes": {"B": 2, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "4795bcdd-6f37-5076-9404-a44be707081c"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/eb36e9422c7cf86d220f49952941a77e94c0646e76a5a65fd4e6674eb6ec7904.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.013635935261845589, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b88e2f5b-0006-5a99-a92d-cf4a8625559e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 7, "uuid": "b88e2f5b-0006-5a99-a92d-cf4a8625559e", "validation": {"axes": {"B": 2, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "b88e2f5b-0006-5a99-a92d-cf4a8625559e"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/8fe56f861959b51196e10bb9b2ae441ac4486fe13724b7812b9dafcb37ad62b5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 2, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.013432308100163937, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e0a836db-4eb5-5404-b90b-feae5f3e075e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 7, "uuid": "e0a836db-4eb5-5404-b90b-feae5f3e075e", "validation": {"axes": {"B": 2, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "e0a836db-4eb5-5404-b90b-feae5f3e075e"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 18, "X": 16, "Y": 10}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/c32d234743cf2934ac3bd116ee7854784e4fbe14abcd4a14b3a03578e2d52226.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "961d1949-75f1-55a6-9950-7c470869589d", "validation": {"axes": {"B": 2, "T": 18, "X": 16, "Y": 10}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.05770906386896968, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "961d1949-75f1-55a6-9950-7c470869589d"}, "variant": "grid_18x10x16", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 18, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/2cace15183b9560b5c3a501b3d67c5f8a8d00e5f581464489250dd92933994f1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "f8bfe021-10d7-522f-9a13-fec1520fd1a9", "validation": {"axes": {"B": 2, "T": 18, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.1499946415424347, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f8bfe021-10d7-522f-9a13-fec1520fd1a9"}, "variant": "grid_18x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 35, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/3f0709e3829436a59369b4e32f6ce8eaa3622408228e5e06dfcb73dab74f5577.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "c33b8ec1-0039-5090-816e-43d244a31e5a", "validation": {"axes": {"B": 2, "T": 35, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.2536184713244438, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c33b8ec1-0039-5090-816e-43d244a31e5a"}, "variant": "grid_35x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 69, "X": 64, "Y": 40}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/66c3a3532532ecd91df77e080a1a0ff9a3f4a5966f54158aa8fd3394bbf9a75d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "6465f201-d904-5e1c-abcb-a2b6b4061ecf", "validation": {"axes": {"B": 2, "T": 69, "X": 64, "Y": 40}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 1.9129024068824947, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6465f201-d904-5e1c-abcb-a2b6b4061ecf"}, "variant": "grid_69x40x64", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 2, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 2, "T": 121, "X": 128, "Y": 80}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/a0e701b7b037fce128e865179f573ae72cb7512b3876a8e598a7cc3a19d324f2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 2, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch2", "requested_outputs": 2, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b2-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b2-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242022.0743153, "step": 0, "uuid": "c6512849-69e2-55f5-8e33-e0f8e148388e", "validation": {"axes": {"B": 2, "T": 121, "X": 128, "Y": 80}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 20.85770556004718, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6512849-69e2-55f5-8e33-e0f8e148388e"}, "variant": "grid_121x80x128", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/c31ed1dd13f35216d39e35f535d31953c2e181ab80b42713acc300d2c3d381ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.878054175991565, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "536d21aa-6cb4-552b-a371-cef96f27abff"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "536d21aa-6cb4-552b-a371-cef96f27abff", "validation": {"axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "536d21aa-6cb4-552b-a371-cef96f27abff"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/adf8b958756b6c5456395607dc101bbf4babcc392c8dddf0daaf5e7988fdc0dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 4.5280489986762404, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b6cfc78f-2787-5033-87b2-88ddfbbd32b1"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "b6cfc78f-2787-5033-87b2-88ddfbbd32b1", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "b6cfc78f-2787-5033-87b2-88ddfbbd32b1"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/37f0ba5b5c1b52e5a97ab7fe3fdcdb561f445d68c22e8f37dc5e7a95ae8f95b3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.41439854772761464, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "32d4a255-e961-5f7b-bcb6-7be48f573e74"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "32d4a255-e961-5f7b-bcb6-7be48f573e74", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "32d4a255-e961-5f7b-bcb6-7be48f573e74"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/364956f348a9133d9142f0255cd16c865a466d06647e5ca28e5d77d78bb8436e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.03254063706845045, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fa5a7eb1-1084-5daa-ab0b-72ac0af9e44f"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "fa5a7eb1-1084-5daa-ab0b-72ac0af9e44f", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "fa5a7eb1-1084-5daa-ab0b-72ac0af9e44f"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f2556bf7ca169063e0a2020c7553acae369f1fa72fdd118c91942ab944a6bd38.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.07182777114212513, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2dcb7d09-5ca5-5b35-aebc-33459f3ca110"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "2dcb7d09-5ca5-5b35-aebc-33459f3ca110", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "2dcb7d09-5ca5-5b35-aebc-33459f3ca110"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/4778b0dea33cad3ceed92ea273af952e12faea6ae601411fb45f9b565e0597e1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03552237572148442, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8828a62a-b05b-5c5f-b48d-46a0dfb68c36"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "8828a62a-b05b-5c5f-b48d-46a0dfb68c36", "validation": {"axes": {"B": 4, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "8828a62a-b05b-5c5f-b48d-46a0dfb68c36"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/635137b3d81cc4dd83db77d34a0a1e41325758e70f42a664078b883ffaca91ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.2588155958801508, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7b693a0c-422c-53af-9ac1-146ee5d09639"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 4, "uuid": "7b693a0c-422c-53af-9ac1-146ee5d09639", "validation": {"axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "7b693a0c-422c-53af-9ac1-146ee5d09639"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/ae747a2ec6853194b019c473eff7fc8d817633dbf01521a280f93371be947dfd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.0028979526832699776, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "abb09f82-ee47-528a-905a-f86b658621f5"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 4, "uuid": "abb09f82-ee47-528a-905a-f86b658621f5", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "abb09f82-ee47-528a-905a-f86b658621f5"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/a31816421975b3858ba9bade08349aae2601b33b97cc1bd98949233a24e74492.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.14615144301205873, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "da56e090-dd99-53a9-84b6-0e98333bab09"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 4, "uuid": "da56e090-dd99-53a9-84b6-0e98333bab09", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "da56e090-dd99-53a9-84b6-0e98333bab09"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/b679d2b17201bb4726e587134eead62f8d43eabdb898fb8590350e83a64b644f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.01586020179092884, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b204c6bd-4c17-553a-b6cd-9b2be4bbae28"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 4, "uuid": "b204c6bd-4c17-553a-b6cd-9b2be4bbae28", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "b204c6bd-4c17-553a-b6cd-9b2be4bbae28"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/fe488dda5f1935cb24923ba16a84a8a2bfd8e09e643d060241852cbe16b343c8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.033811728935688734, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "05a21a06-56d8-54e4-b6f8-18ee1a3235fd"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 4, "uuid": "05a21a06-56d8-54e4-b6f8-18ee1a3235fd", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "05a21a06-56d8-54e4-b6f8-18ee1a3235fd"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/336d298af4065f77b2745ad627b16cecfc210961e7cd7f03e7811329fc46e985.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03289396455511451, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e18b75bf-014a-57d6-a428-42f15d177aa3"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 4, "uuid": "e18b75bf-014a-57d6-a428-42f15d177aa3", "validation": {"axes": {"B": 4, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "e18b75bf-014a-57d6-a428-42f15d177aa3"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/be4426ac7c7f9d97045c85b946f0f0db830b0bfebe7e99e64e695fbdbdc775cd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.23946930607780814, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "991c130c-f455-5544-8774-7a0bcb71ca17"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 7, "uuid": "991c130c-f455-5544-8774-7a0bcb71ca17", "validation": {"axes": {"B": 4, "Nk": 2352, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "991c130c-f455-5544-8774-7a0bcb71ca17"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/8a570e96f2f6df0c84fac6d82c51255a168059ba49db1b45b340814170d0a78a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0028513921424746513, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "86da2b5c-41a3-5409-8bde-0c0b85162668"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 7, "uuid": "86da2b5c-41a3-5409-8bde-0c0b85162668", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "86da2b5c-41a3-5409-8bde-0c0b85162668"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/66e8f5b1ce60519a1ae900eb57ebd72b71570d435c94613c891887437b8b4f6a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.10912138083949685, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6cb3617c-6d8b-5c31-bf25-d5220009223e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 7, "uuid": "6cb3617c-6d8b-5c31-bf25-d5220009223e", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "6cb3617c-6d8b-5c31-bf25-d5220009223e"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/36f7fdef855507afa6a5961da446766b60711fa0692e01537ce2f1399341f027.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.01061081001535058, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e28ce398-4258-5170-a14f-a247318ff3d3"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 7, "uuid": "e28ce398-4258-5170-a14f-a247318ff3d3", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "e28ce398-4258-5170-a14f-a247318ff3d3"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 2352}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/950be58dd29df63051503b9927fc8f85a8f22aac673e06098559146c44bba348.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.020939147099852562, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "efa942e5-4b8c-5cb5-9554-232fa3b7b889"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 7, "uuid": "efa942e5-4b8c-5cb5-9554-232fa3b7b889", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 2352}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "efa942e5-4b8c-5cb5-9554-232fa3b7b889"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2352, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/a128004894205e6e95fe9589418209885a17a4e8c4cd31e3c374edb44fd1352b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.020583586767315865, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d3cdef82-ba3d-5778-9b1b-81864342cfa3"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 7, "uuid": "d3cdef82-ba3d-5778-9b1b-81864342cfa3", "validation": {"axes": {"B": 4, "Nk": 2352, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "d3cdef82-ba3d-5778-9b1b-81864342cfa3"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 9, "X": 24, "Y": 14}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/076a2eff76e16fa504d9f089868859adc02144bbac05eb521cbd17fe55174854.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "2882e24d-2dc3-5644-8e69-8c22c2b13dda", "validation": {"axes": {"B": 4, "T": 9, "X": 24, "Y": 14}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.11728813499212265, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2882e24d-2dc3-5644-8e69-8c22c2b13dda"}, "variant": "grid_9x14x24", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 9, "X": 48, "Y": 28}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/6e0a1c765dff714d2fc84ac85797f5885509c16b48b1d1e054aa9f456164d1df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "7a7c903b-01cd-5e52-a7ad-610af7f4b4d4", "validation": {"axes": {"B": 4, "T": 9, "X": 48, "Y": 28}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.7234842558391392, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7a7c903b-01cd-5e52-a7ad-610af7f4b4d4"}, "variant": "grid_9x28x48", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 17, "X": 48, "Y": 28}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/6611bff57cf8e17b278e8a3d1312e8e1b35d316dd54c42fffbe05325d775133e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "14c7aef4-8e42-5a1b-8f42-16146a940901", "validation": {"axes": {"B": 4, "T": 17, "X": 48, "Y": 28}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.30765239894390106, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "14c7aef4-8e42-5a1b-8f42-16146a940901"}, "variant": "grid_17x28x48", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 33, "X": 96, "Y": 56}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/e082ee0b4559b74142ece47d8528fed46c8df3761c0f3f1ef2b37d6a87de9b48.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "0253ef4d-943e-5839-9cb0-8d384be40c67", "validation": {"axes": {"B": 4, "T": 33, "X": 96, "Y": 56}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 2.1867749830707908, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0253ef4d-943e-5839-9cb0-8d384be40c67"}, "variant": "grid_33x56x96", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 49, "X": 192, "Y": 112}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/ea644641dafb04a96f9677985426305371b04a45999ec0d938d5eb5193c350dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A red ceramic teapot on a wooden table beside a window. The camera slowly moves closer as steam rises from the spout. Soft afternoon light and gentle room ambience.", "seed": 42, "size": "768x448"}, "request_id": "ltx25_text_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242247.1026254, "step": 0, "uuid": "d115c6f3-2ee4-5393-affc-0f41a03e63aa", "validation": {"axes": {"B": 4, "T": 49, "X": 192, "Y": 112}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 35.441113733220845, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d115c6f3-2ee4-5393-affc-0f41a03e63aa"}, "variant": "grid_49x112x192", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/73659cd428a2c0d0fb25ae46514c9697d1ac106f6c76e6d411054e060ac19c72.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.17746720975264907, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d3373710-c368-59f3-9515-7b961b87d57c"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "d3373710-c368-59f3-9515-7b961b87d57c", "validation": {"axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "d3373710-c368-59f3-9515-7b961b87d57c"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/adf8b958756b6c5456395607dc101bbf4babcc392c8dddf0daaf5e7988fdc0dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.002828259952366352, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0dedb1b6-22f5-5317-8887-ca42e277b7c2"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "0dedb1b6-22f5-5317-8887-ca42e277b7c2", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "0dedb1b6-22f5-5317-8887-ca42e277b7c2"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/ac9cfc44c2fc20a87353a40849946e0d5a39dcfd0d04fa74370c2658532ea341.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.062051201704889536, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4cdcb87c-f1a0-5604-887b-c7a07ba2e4c4"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "4cdcb87c-f1a0-5604-887b-c7a07ba2e4c4", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "4cdcb87c-f1a0-5604-887b-c7a07ba2e4c4"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/e1012492f0b34016af3e1acc33f91e691f25bacb5909e985b1e61f5288922b1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.01036933995783329, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a4fe2d19-6cec-564d-a702-0312dfa23b28"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "a4fe2d19-6cec-564d-a702-0312dfa23b28", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "a4fe2d19-6cec-564d-a702-0312dfa23b28"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/34fd06382976c9e5ea5abbfa690437b3e6f0a00cb34b1619806ab7e8d208dd59.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.02837013592943549, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d74d520c-1fff-5457-93d0-0301994b6f31"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "d74d520c-1fff-5457-93d0-0301994b6f31", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "d74d520c-1fff-5457-93d0-0301994b6f31"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/0bc98e2efccf97fdf3a051c8d2c7fae2880a72374422c82d65d5f14a5ff1eea2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.011355063412338495, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2f44bd82-037b-5ad2-bc34-88f523346678"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "2f44bd82-037b-5ad2-bc34-88f523346678", "validation": {"axes": {"B": 4, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "2f44bd82-037b-5ad2-bc34-88f523346678"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/81e125e358f1b6d1bde96d3fe924fffe68bdadfa8dcccbde6573dea653260b0f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06526764808222651, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3f618d6d-4678-54d9-a3ef-3b69bb773f14"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 4, "uuid": "3f618d6d-4678-54d9-a3ef-3b69bb773f14", "validation": {"axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "3f618d6d-4678-54d9-a3ef-3b69bb773f14"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/be694ffea3a4793cc89856d6ffaebe59c32ca2d595c3a38ac2e74ee3315894c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0028808768838644028, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c2e44540-2f70-5245-a34c-2b91bf89a7ae"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 4, "uuid": "c2e44540-2f70-5245-a34c-2b91bf89a7ae", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "c2e44540-2f70-5245-a34c-2b91bf89a7ae"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/44165e2052f3530b8661169e52cac279c9f4f97a07434baba270b19941bdfed6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06193388858810067, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f08aaae1-1cb4-5945-8f8f-00afb03f90a1"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 4, "uuid": "f08aaae1-1cb4-5945-8f8f-00afb03f90a1", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "f08aaae1-1cb4-5945-8f8f-00afb03f90a1"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/dc3c34114e51602b6b757e556412e1c7f2eb629258889edde6b2e93f6d6d0e0f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.010333027224987745, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9109ad43-fa43-5892-a18b-3351bd60ede1"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 4, "uuid": "9109ad43-fa43-5892-a18b-3351bd60ede1", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "9109ad43-fa43-5892-a18b-3351bd60ede1"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/7fb91da18ba2b30f4fdd0dd7d8993549d98e86f7671ce9c5e36a5dd3ec22d910.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.011175353080034256, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c1cd8d74-fd40-5256-a30b-5e1cb4ef5c52"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 4, "uuid": "c1cd8d74-fd40-5256-a30b-5e1cb4ef5c52", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "c1cd8d74-fd40-5256-a30b-5e1cb4ef5c52"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/6ed79d862da987f131ea8a0b6cf38c3d784a77baac69e2838e30eb95753559ea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.011115758214145899, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ee702d9a-f86d-5b59-a5b0-e6b421997b87"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 4, "uuid": "ee702d9a-f86d-5b59-a5b0-e6b421997b87", "validation": {"axes": {"B": 4, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "ee702d9a-f86d-5b59-a5b0-e6b421997b87"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/1dc486c0c8c615fa83607e7d74c6ce1c502d7c9059cc3a9d73ff93dd76936d9b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0625], "seconds": 0.06480937078595161, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "080b6a8e-f168-5131-9f1f-a4eeaf37cbbf"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 7, "uuid": "080b6a8e-f168-5131-9f1f-a4eeaf37cbbf", "validation": {"axes": {"B": 4, "Nk": 1120, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0625], "status": "passed", "uuid": "080b6a8e-f168-5131-9f1f-a4eeaf37cbbf"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/47887c9822e410c699f8ccbb75e31b76215477a9792609d9cfca9f2b98c9b323.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.002844560891389847, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1e22f1b7-23e2-5dee-92d5-f74e6c6ca3ea"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 7, "uuid": "1e22f1b7-23e2-5dee-92d5-f74e6c6ca3ea", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "1e22f1b7-23e2-5dee-92d5-f74e6c6ca3ea"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/bb81a841233c5f47be77ce44b26aa95bd0786b68874539cd78e3e16190b248b2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06066583702340722, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e58ece45-84b5-50dd-95b8-2becd9cb9d3d"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 7, "uuid": "e58ece45-84b5-50dd-95b8-2becd9cb9d3d", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "e58ece45-84b5-50dd-95b8-2becd9cb9d3d"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/27f9e4d855cc6eb836dba6d93c966cdfc3843f309ec845aad0484d684bb100dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.010276524815708399, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e9ad38c8-2a06-58d9-b32a-a96e3255e630"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 7, "uuid": "e9ad38c8-2a06-58d9-b32a-a96e3255e630", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0009765625], "status": "passed", "uuid": "e9ad38c8-2a06-58d9-b32a-a96e3255e630"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 51, "Nq": 1120}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/1f0e5f2116b8b9bf8def3832fe6f10d9acab52d71dbc59864a45d7f63cb26ba7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.010753185953944921, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "645dc728-94d6-5801-a837-9a1f4201b237"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 7, "uuid": "645dc728-94d6-5801-a837-9a1f4201b237", "validation": {"axes": {"B": 4, "Nk": 51, "Nq": 1120}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "645dc728-94d6-5801-a837-9a1f4201b237"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1120, "Nq": 51}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/abc7cb1fb3f8f8c4595959440175a159e57211d179c005865e285de050a398aa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.010692563839256763, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6fccac8a-d4ec-54e4-99d1-48ded5d20120"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 7, "uuid": "6fccac8a-d4ec-54e4-99d1-48ded5d20120", "validation": {"axes": {"B": 4, "Nk": 1120, "Nq": 51}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "6fccac8a-d4ec-54e4-99d1-48ded5d20120"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 9, "X": 16, "Y": 10}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/058d7df956e2835de0d256063d5e140b40ee61d7454572db692f5b8203e316c0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "2f58848f-7e64-55b7-ad2d-5de5d12d6373", "validation": {"axes": {"B": 4, "T": 9, "X": 16, "Y": 10}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.05699899513274431, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2f58848f-7e64-55b7-ad2d-5de5d12d6373"}, "variant": "grid_9x10x16", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 9, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/fccb4672737c62149a905287677b6ca1f8cb82e5c8683e21e4d863e50fbe3929.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "328bbc42-0b45-5dec-bb11-0a53ec0f4ea5", "validation": {"axes": {"B": 4, "T": 9, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.11679140664637089, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "328bbc42-0b45-5dec-bb11-0a53ec0f4ea5"}, "variant": "grid_9x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 17, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/27dd5e1537155a92dbce19158a59804f1a3808e19d7d8730cc3fb15573361683.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "ece8abe8-c6b6-55f8-8585-e455a076854b", "validation": {"axes": {"B": 4, "T": 17, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.14409761037677526, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ece8abe8-c6b6-55f8-8585-e455a076854b"}, "variant": "grid_17x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 33, "X": 64, "Y": 40}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/46aec19ef278cd6f62346d49185e694f8b1b21ea80d43fc9961ee799ae4ea9ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "abe8cf54-16a2-5f37-b119-71880cdf77d5", "validation": {"axes": {"B": 4, "T": 33, "X": 64, "Y": 40}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 1.0312290657311678, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "abe8cf54-16a2-5f37-b119-71880cdf77d5"}, "variant": "grid_33x40x64", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 49, "X": 128, "Y": 80}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/7413f14c268624ac29d9d00451d1d41936027ecdaf67bb7ec5084aaf51199108.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "reference_image_sha256": "e6f2095f119b4ca59cace7b8969cac987408fb38d80a76e68ffff043a71ecc80", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "input_reference": "/task/raw/image-b1-r1/results/square.png", "num_frames": 49, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "The camera slowly moves around the ceramic teapot. Soft steam rises from its spout while the table and window remain still. Quiet indoor ambience.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_image_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242361.3072696, "step": 0, "uuid": "6da01593-e133-55fe-8eb8-69be7217be58", "validation": {"axes": {"B": 4, "T": 49, "X": 128, "Y": 80}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 16.0095423348248, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6da01593-e133-55fe-8eb8-69be7217be58"}, "variant": "grid_49x80x128", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/a590229d40d8febdc16a1eaa393696a62fe4fa658ec4dc5a94c46d5a18b0ab46.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.2070871451869607, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "33f396cc-71fe-5f37-961b-20b86a8c078e"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "33f396cc-71fe-5f37-961b-20b86a8c078e", "validation": {"axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "33f396cc-71fe-5f37-961b-20b86a8c078e"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/fecadb73a7d45ca1ad72ae112a165ff81c539bed4b65754bd2e70b3fae36792a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.008462416008114815, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "29041579-efc8-573b-8642-1ea8af6a2cdf"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "29041579-efc8-573b-8642-1ea8af6a2cdf", "validation": {"axes": {"B": 4, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "29041579-efc8-573b-8642-1ea8af6a2cdf"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/c64f2f27fbfdb4e9633ba25210464ca7caed3aafa02a5d722aca732e432bf272.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.10490890173241496, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ca03b08-4582-5408-8c48-6f4b30225045"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "6ca03b08-4582-5408-8c48-6f4b30225045", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "6ca03b08-4582-5408-8c48-6f4b30225045"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/968f86a7091fb914372095ad6b33d130fb2987022c3f2c93c172a8c25c9a5653.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.015299574937671423, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fbb0bb64-f32b-589b-bc7f-48b571a90daf"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "fbb0bb64-f32b-589b-bc7f-48b571a90daf", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "fbb0bb64-f32b-589b-bc7f-48b571a90daf"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/6d47140ef82be5b88da342531918c9c66723d6423fb17ed2280ccf8ccc69a7ad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.030222739558666945, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4e69b9b6-fb8d-51c1-b88d-d0b346d8bf92"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "4e69b9b6-fb8d-51c1-b88d-d0b346d8bf92", "validation": {"axes": {"B": 4, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.001953125], "status": "passed", "uuid": "4e69b9b6-fb8d-51c1-b88d-d0b346d8bf92"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/3839b8c636327b9999874fb4fa41e0dfd47e64a8ffb9ae2b63d0a09d89b7deb8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.028602121863514185, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2b784abf-d74d-59d2-840f-d750c00907cd"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "2b784abf-d74d-59d2-840f-d750c00907cd", "validation": {"axes": {"B": 4, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "2b784abf-d74d-59d2-840f-d750c00907cd"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/e3cfeafe29ba4ccc424673195bff1e85a8605d55d755c2aa94459fbfd2c9d53d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.20110497903078794, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "919246a0-b394-5215-b9c0-afda319ecefa"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 4, "uuid": "919246a0-b394-5215-b9c0-afda319ecefa", "validation": {"axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "919246a0-b394-5215-b9c0-afda319ecefa"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/7dfffb29789c57c5c833f443bcaaaccad9577b37a83f9768de4811f5fe088b9c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004404451232403517, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b19d6db5-9600-536d-8f4c-543dba8b95d0"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 4, "uuid": "b19d6db5-9600-536d-8f4c-543dba8b95d0", "validation": {"axes": {"B": 4, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "b19d6db5-9600-536d-8f4c-543dba8b95d0"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/e357b3d87f94327c5b844eb98c497525c6e86c0d5b5eab9f7cb28987f16aa86a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.10427208803594112, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f0dfee8e-8c32-5004-a09c-248d46e01062"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 4, "uuid": "f0dfee8e-8c32-5004-a09c-248d46e01062", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "f0dfee8e-8c32-5004-a09c-248d46e01062"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/359b3cd62377303671ce84039f6fbeb1fee2a37111d95d4ca0f9556c6b6f2ad6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.013297016266733408, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a5cb657f-b218-5f59-b04e-23be91dee147"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 4, "uuid": "a5cb657f-b218-5f59-b04e-23be91dee147", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "a5cb657f-b218-5f59-b04e-23be91dee147"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/56bb13dd44f373a0b5d110ec76a266f9db188dae26525813f37193ab71828881.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.028794713784009218, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ac66abe-111f-58be-8c6f-51b54e0e7336"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 4, "uuid": "6ac66abe-111f-58be-8c6f-51b54e0e7336", "validation": {"axes": {"B": 4, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "6ac66abe-111f-58be-8c6f-51b54e0e7336"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/a0e918cdaa198fe06cac9bd2c8ef76648333620ec31b75131d43497efe6c5f88.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.02881968440487981, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "365db76e-b098-5d94-84a6-2e0791f4ea86"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 4, "uuid": "365db76e-b098-5d94-84a6-2e0791f4ea86", "validation": {"axes": {"B": 4, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "365db76e-b098-5d94-84a6-2e0791f4ea86"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/4ecff6d01b884e4bde1f794acee60c5ccc8c1bac3500513e791ade113f655a67.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0625], "seconds": 0.2003156109713018, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2994bf07-9d15-5981-9289-2340dfb17b32"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 7, "uuid": "2994bf07-9d15-5981-9289-2340dfb17b32", "validation": {"axes": {"B": 4, "Nk": 2560, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0625], "status": "passed", "uuid": "2994bf07-9d15-5981-9289-2340dfb17b32"}, "variant": "attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 126, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/b8a468093af919ab0a59b5b41c87d3d8dffd8370b3bd636a498b5f6eee148f4f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004348859190940857, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a656cf2a-7244-59dd-8937-83cc198c0202"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn1", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 7, "uuid": "a656cf2a-7244-59dd-8937-83cc198c0202", "validation": {"axes": {"B": 4, "Nk": 126, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.015625], "status": "passed", "uuid": "a656cf2a-7244-59dd-8937-83cc198c0202"}, "variant": "audio_attn1", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d128", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d128/b2f79fd785485bc4a70627c7440b6b61af74f412d24208756f12511d45787b2f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.10353726986795664, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "637331d6-ea8d-5f4d-8a22-d4135f99cf2f"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 7, "uuid": "637331d6-ea8d-5f4d-8a22-d4135f99cf2f", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d128", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "637331d6-ea8d-5f4d-8a22-d4135f99cf2f"}, "variant": "attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 1024, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/f1c47c5260c2a624ab4d764d45c4d3fb29499b1d588b1af1d2004eed4da59ea9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.01301031420007348, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd1288ed-8967-57c0-8c57-a6a5f6194392"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_attn2", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 7, "uuid": "bd1288ed-8967-57c0-8c57-a6a5f6194392", "validation": {"axes": {"B": 4, "Nk": 1024, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.00390625], "status": "passed", "uuid": "bd1288ed-8967-57c0-8c57-a6a5f6194392"}, "variant": "audio_attn2", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 126, "Nq": 2560}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/b23182a122264c0f9362b803e9a72c25f11d1bd4f8ab957c2451fe3a278ac6cb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.02794422162696719, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0f7dec0c-09a6-5895-ad37-d817ddc5ea66"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.audio_to_video_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 7, "uuid": "0f7dec0c-09a6-5895-ad37-d817ddc5ea66", "validation": {"axes": {"B": 4, "Nk": 126, "Nq": 2560}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.0078125], "status": "passed", "uuid": "0f7dec0c-09a6-5895-ad37-d817ddc5ea66"}, "variant": "audio_to_video_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Preserve post-normalization/RoPE Q/K/V; expand the engine visibility mask without changing it.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "Nk": 2560, "Nq": 126}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "noncausal_dense_attention_h32_d64", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/noncausal_dense_attention_h32_d64/1b40031b1d965be1c6ec6b76e031e214bcf845fe43936dc04ed17e422590d0db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "interface_curation": {"input_values_and_engine_output_unchanged": true, "invariants_verified_for_all_workloads": true, "previous_replay": {"axes": {"B": 4, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.026809928938746452, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6fde8bd-ad59-57be-bd10-75481240b2a9"}, "removed_capture_only_inputs": ["sm_scale"], "replay_date": "2026-10-19"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "transformer_blocks.0.video_to_audio_attn", "operation": "LTX2Attention.attn", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "sdpa_backend": "math", "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 7, "uuid": "c6fde8bd-ad59-57be-bd10-75481240b2a9", "validation": {"axes": {"B": 4, "Nk": 2560, "Nq": 126}, "definition": "noncausal_dense_attention_h32_d64", "max_absolute_errors": [0.03125], "status": "passed", "uuid": "c6fde8bd-ad59-57be-bd10-75481240b2a9"}, "variant": "video_to_audio_attn", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 18, "X": 16, "Y": 10}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h32_d64_kt3_ky7_kx7/462e74c323d7587b91648e116007e772d5fbd962280981391b58e1b9989dfb9e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h32_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "867a9c5a-089e-5005-bb09-db1354c3e298", "validation": {"axes": {"B": 4, "T": 18, "X": 16, "Y": 10}, "definition": "neighborhood_attention_3d_h32_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.11142154317349195, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "867a9c5a-089e-5005-bb09-db1354c3e298"}, "variant": "grid_18x10x16", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 18, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h16_d64_kt3_ky7_kx7/af373e4f2d1695d315be0b58e993105ee1624be85e6a284e893fd116b9e9e8df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h16_3x7x7", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "e8f46b73-bd40-55e5-96d0-32dc0199b4f1", "validation": {"axes": {"B": 4, "T": 18, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h16_d64_kt3_ky7_kx7", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.2318596919067204, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e8f46b73-bd40-55e5-96d0-32dc0199b4f1"}, "variant": "grid_18x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 35, "X": 32, "Y": 20}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/205af6e016173fee56a83608a06e25d1799905a1e18a1f000946968143f1431b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "9866c54d-a157-5fdd-89d1-84af0fb714e6", "validation": {"axes": {"B": 4, "T": 35, "X": 32, "Y": 20}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.2929818043485284, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9866c54d-a157-5fdd-89d1-84af0fb714e6"}, "variant": "grid_35x20x32", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 69, "X": 64, "Y": 40}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h8_d64_kt3_ky5_kx5/556acca09dd870fc2ff8429c35d61ae784b0fe36ae61c7fbe2acb1381971393a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h8_3x5x5", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "e575ca18-1da8-5e40-8439-66d5e614c4c5", "validation": {"axes": {"B": 4, "T": 69, "X": 64, "Y": 40}, "definition": "neighborhood_attention_3d_h8_d64_kt3_ky5_kx5", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 2.1642743698321283, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e575ca18-1da8-5e40-8439-66d5e614c4c5"}, "variant": "grid_69x40x64", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_attention_batch": 4, "adapter": "Capture the complete grid after normalization, query scaling and RoPE, before output projection.", "audio_precision_adapter": "Select PyTorch math SDPA for every dimension-64 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "audio_sdpa_backend": "math", "axes": {"B": 4, "T": 121, "X": 128, "Y": 80}, "capture_steps": [0, 4, 7], "conversion": {"conv_vae_distribution": "comfyicu/LTX-2.5", "conv_vae_revision": "1026e00f808dbd9601911479b22d65b760737dc8", "converter_revision": "efe46a5b177af67c64b3666da8ddddac7ee12163", "original_model": "Lightricks/LTX-2.5", "original_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "public_distribution": "pinecoresystems/LTX-2.5-serving", "revision": "93deb76408feee57d04b3cfeed3ded56a7a068ee", "source_sha256": {"diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors": "31eb3cad89b9e54e99dd3baf286f70825ac4f6c660a70d9184d895be76d7bff4", "text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors": "ef7243612fdae7a75cb4d5cee9433e81380675fb6c213bd98ae74a9cd16561d1", "vae/ltx-2.5-audio-vae-bf16.safetensors": "c52733d37f6a7fb7949c3dc0fb468c6cb2169e4d836983a73babb9f0d54837a5", "vae/ltx-2.5-video-vae-bf16.safetensors": "847e14ca7f3355debca0cea4eaa24ac0fbcdf0061da054ac89ca638a869ddba3", "vae/ltx-2.5-video-vae-conv-bf16.safetensors": "685b06ee3d9b2039647698fc4ea33175112462fc374e2777312c907897dfce8d"}, "text_encoder_adapter": "Lossless parameter-name mapping, including vision/audio embeddings and tied LM head.", "validation": "Every component loaded with strict=True; all source file SHA256 values verified."}, "decoder_adapter": "Retain the original optional keyframe tag; use the complete ordinary-video decoder path.", "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/neighborhood_attention_3d_h4_d64_kt11_ky11_kx11/d63cf29cf8ca873c1cc6990699c491d629a234cbffb27a612f984490395300e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "model": "Lightricks/LTX-2.5", "model_revision": "49b093a39c19169a8a81721dcc541e07e945232f", "module_path": "diffusion_decoder.attention_h4_11x11x11", "operation": "LTX2VideoVaeNeighborhoodAttention", "request": {"enhance_prompt": false, "extra_body": {"use_diffusion_decoder": true}, "fps": 24, "guidance_scale": 1.0, "num_frames": 121, "num_inference_steps": 8, "num_outputs_per_prompt": 4, "prompt": "A small wooden boat crosses a calm lake at sunrise. The camera follows smoothly as ripples spread behind the boat. Birds call softly from the shore.", "seed": 42, "size": "512x320"}, "request_id": "ltx25_long_video_batch4", "requested_outputs": 4, "sampling_sigmas": [1, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875], "server_command": ["sglang", "serve", "--model-path", "/task/models/ltx25", "--model-id", "LTX-2.5-Diffusers", "--host", "127.0.0.1", "--port", "30400", "--performance-mode", "speed", "--scheduler-port", "30420", "--master-port", "30421", "--attention-backend", "torch_sdpa", "--enable-torch-compile", "false", "--enable-breakable-cuda-graph", "false", "--warmup-mode", "off", "--load-diffusion-decoder", "true", "--vae-tiling", "false", "--vae-sp", "false", "--input-save-path", "/task/raw/ltx-b4-r4/results/inputs", "--diffusion-decoder-tiling", "false", "--output-path", "/task/raw/ltx-b4-r4/results/videos"], "source": "https://github.com/sgl-project/sglang/tree/684fea8a/python/sglang/multimodal_gen", "started_at": 1792242401.3913639, "step": 0, "uuid": "b188ec47-6dff-5a21-a6fb-1f0d4431ee7a", "validation": {"axes": {"B": 4, "T": 121, "X": 128, "Y": 80}, "definition": "neighborhood_attention_3d_h4_d64_kt11_ky11_kx11", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 39.63709596823901, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b188ec47-6dff-5a21-a6fb-1f0d4431ee7a"}, "variant": "grid_121x80x128", "video_precision_adapter": "Select PyTorch math SDPA for every dimension-128 transformer attention call; keep BF16 inputs/outputs and FP32 intermediates; normalize output storage to the engine's sequence-major flattening contract.", "video_sdpa_backend": "math"}
{"actual_requested_output_tokens": 107, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 2, "NP": 2, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f85737bdc0d96ac0bc073f36bdc22f7a459f9531c81e826431d9d26cc9974ef9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f85737bdc0d96ac0bc073f36bdc22f7a459f9531c81e826431d9d26cc9974ef9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1cef83d61ca56b0bacc92e5c55626ed3010575136bc32e54db7db753ea383488.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d52bde5a8d11428bb129ece05fc0404b150685d4ecad5d2cc1d6d9efb0779802.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [129], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "691b7d0c-2cf5-57f0-8008-6c37d24d2ced", "validation": {"axes": {"B": 1, "MP": 2, "NP": 2, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.49019037326797843, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "691b7d0c-2cf5-57f0-8008-6c37d24d2ced"}}
{"actual_requested_output_tokens": 107, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f59183baa1c3a496bfa8fc74042c98152632cd5f5b00f351494abd530499419c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [129], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "d15d9f9b-cd28-5c5f-aea1-a0b1c33d63e6", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.12573457602411509, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d15d9f9b-cd28-5c5f-aea1-a0b1c33d63e6"}}
{"actual_requested_output_tokens": 107, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 2, "NP": 2, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f85737bdc0d96ac0bc073f36bdc22f7a459f9531c81e826431d9d26cc9974ef9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f85737bdc0d96ac0bc073f36bdc22f7a459f9531c81e826431d9d26cc9974ef9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1cef83d61ca56b0bacc92e5c55626ed3010575136bc32e54db7db753ea383488.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5cc2958486887b349321f89b318f701f38ac21ed7ceae16a12e92f691d904a3c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [130], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "43f45cd3-694c-5186-a45e-16e2afec402a", "validation": {"axes": {"B": 1, "MP": 2, "NP": 2, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016061612404882908, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "43f45cd3-694c-5186-a45e-16e2afec402a"}}
{"actual_requested_output_tokens": 107, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/05037ac00f190324af5c8c833e4c38a306a2a8360dacccc6f105d6ca41113ede.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [130], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "974cb7c4-115c-5c64-8658-6cddb2098495", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0026712389662861824, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "974cb7c4-115c-5c64-8658-6cddb2098495"}}
{"actual_requested_output_tokens": 107, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 2, "NP": 2, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f85737bdc0d96ac0bc073f36bdc22f7a459f9531c81e826431d9d26cc9974ef9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f85737bdc0d96ac0bc073f36bdc22f7a459f9531c81e826431d9d26cc9974ef9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1cef83d61ca56b0bacc92e5c55626ed3010575136bc32e54db7db753ea383488.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d5f3117c5d5df4cfd0f2c7f8bcddd7d470a2735f342119a109647f51d6ef2e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [131], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "a81fb12c-a07e-58cf-b094-a6f7120c35da", "validation": {"axes": {"B": 1, "MP": 2, "NP": 2, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013757487758994102, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a81fb12c-a07e-58cf-b094-a6f7120c35da"}}
{"actual_requested_output_tokens": 107, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c1141f239589eded86b810412dd53a9fc8bf86b07cdb9cc96ffb5a9ebb577e09.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [131], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "04fec133-17a6-5948-ac11-807c4063a375", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0024944180622696877, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "04fec133-17a6-5948-ac11-807c4063a375"}}
{"actual_requested_output_tokens": 107, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f646dc5e8111da696ef3c88dab604607b387169711922b98646618291a9abf9c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f646dc5e8111da696ef3c88dab604607b387169711922b98646618291a9abf9c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6fbbf18c2ad210477d572f887d0fbbd1669bcd3a76aec4be709243ab74112efa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [132], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "1423e2b8-bc82-5449-a246-167374efa8c0", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001559934113174677, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1423e2b8-bc82-5449-a246-167374efa8c0"}}
{"actual_requested_output_tokens": 107, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0813c5520d71693f848474a0828cddb4152180945c2156353f997b7ef5d1cc0c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [132], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "80a15e07-f18a-505d-87fe-7ca5a6c99f71", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0025742421858012676, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "80a15e07-f18a-505d-87fe-7ca5a6c99f71"}}
{"actual_requested_output_tokens": 107, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/be9c626839a20f7a98565355396d1fe19604170962762f604f6a110cd57e1533.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/be9c626839a20f7a98565355396d1fe19604170962762f604f6a110cd57e1533.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7b45a51cb4043ab43d747922f07950e7f1e0f04833c83970f2251d7603e22865.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [181], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 52, "uuid": "655d4a1e-549d-5c7d-a427-e408e1474695", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0014491910114884377, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "655d4a1e-549d-5c7d-a427-e408e1474695"}}
{"actual_requested_output_tokens": 107, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9ec77c067392c5bc68a85da18e96db54b85df55398aa8b643b85bf7f2dd01df1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4d8c659ebc7c8ea311eb7e327d2047320b4e614bdbf74b067ee4fada5a550a30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [181], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 52, "uuid": "0e5efda1-5d5c-53bb-8907-47da5d594ce5", "validation": {"axes": {"B": 1, "MP": 3, "NP": 3, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0025380710139870644, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0e5efda1-5d5c-53bb-8907-47da5d594ce5"}}
{"actual_requested_output_tokens": 107, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 4, "NP": 4, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6dd9c1b483112142861f25f31da4aa9195c1b7f94cd60b8c2a7bc71f487266eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6dd9c1b483112142861f25f31da4aa9195c1b7f94cd60b8c2a7bc71f487266eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/15e82920c067b9d27b736ae594b93d44dc46d32721fdff59960854204e798aac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6e5e375900ed89c24438be98fe94d9b228d98a0666147088a37d653dcab257ef.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [234], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 105, "uuid": "d00c9ed0-3109-5bf2-a233-046f04e83509", "validation": {"axes": {"B": 1, "MP": 4, "NP": 4, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0015910831280052662, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d00c9ed0-3109-5bf2-a233-046f04e83509"}}
{"actual_requested_output_tokens": 107, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 4, "NP": 4, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b5a95c9e043e082136127b7916fac4c2882a2f6e1806aaca8914ed4801c4638a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "03674a4fc8fea34b383a3015f41b09b0c441f14bdbc25e218c1305307d6f2f08", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/15e82920c067b9d27b736ae594b93d44dc46d32721fdff59960854204e798aac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [234], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 105, "uuid": "a3c3d1cf-9da2-55cb-9148-b0a8ced14e20", "validation": {"axes": {"B": 1, "MP": 4, "NP": 4, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0026835682801902294, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a3c3d1cf-9da2-55cb-9148-b0a8ced14e20"}}
{"actual_requested_output_tokens": 1813, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 46, "NP": 46, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/20f44cc69fb161f520059a7d0a2b6f3fa3dc16d31790f0a24fb8da8f32f1b063.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/710f365f32e6c03b0438739278a5bc8f7e0ccc32a6a5257f0e949ba2fac8e81c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8d442947e6b7b0460308ee9d6a914c663e1e48cae9aea60cfac28ec56a24ca4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/53dadedc23214e11ec438e1e26a513e12a331b7280d1879289454603a9ebd057.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7b70c914-edb9-5caa-be9e-5adba28327d0", "validation": {"axes": {"B": 1, "MP": 46, "NP": 46, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016893609426915646, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7b70c914-edb9-5caa-be9e-5adba28327d0"}}
{"actual_requested_output_tokens": 1813, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/82fe9efd5392f21da0e52c6806af889bca557c96157689e513c03d0b6574b06e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/01233d2f05f6d6cfd452a92e97d340101e2cd1993071735db4dc568625dcf9c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "98b37bba-29c7-56f9-8929-57e172095f9c", "validation": {"axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004171859938651323, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "98b37bba-29c7-56f9-8929-57e172095f9c"}}
{"actual_requested_output_tokens": 1813, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 46, "NP": 46, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a97ea22d5d4f63ba629f41c2f225f8bbc4b8afbcdfe8d4cd3da5f99b2ceaf0a6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7c69778e10e85779569fdceefbcdadb26d3109516608ad66535d4e1d17690dee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8d442947e6b7b0460308ee9d6a914c663e1e48cae9aea60cfac28ec56a24ca4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a68dc84581baa94785572a07b0210280689d7fc9365e1830b3720cfcda0eb4f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "95127ae0-3d44-5efd-973f-fe4274165438", "validation": {"axes": {"B": 1, "MP": 46, "NP": 46, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001458143349736929, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "95127ae0-3d44-5efd-973f-fe4274165438"}}
{"actual_requested_output_tokens": 1813, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/402f075761ebf3f333ebb5946af4b4fc46e24d9b17e935df2ee3ef45252983cb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/01233d2f05f6d6cfd452a92e97d340101e2cd1993071735db4dc568625dcf9c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "dccb8d35-e3c0-592b-8037-64908da3cec7", "validation": {"axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004039887338876724, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dccb8d35-e3c0-592b-8037-64908da3cec7"}}
{"actual_requested_output_tokens": 1813, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 46, "NP": 46, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5a6f58b5e4efbb8211cd7596d2b88283ec6818cb03c631fbcfb77815224c65b2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ce3aa9a936109cd87e5352a28f5710b1601095129a2b00f62790dcb365dfb839.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8d442947e6b7b0460308ee9d6a914c663e1e48cae9aea60cfac28ec56a24ca4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9e16cc9595d4d8bcbf09cd9c69b6d0fe6056461354bf7e38b300554fc665ee41.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "5c45f586-4d11-577e-a101-5b7bf8deca7d", "validation": {"axes": {"B": 1, "MP": 46, "NP": 46, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0014333920553326607, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5c45f586-4d11-577e-a101-5b7bf8deca7d"}}
{"actual_requested_output_tokens": 1813, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/879b1729279eb373091d2d3c3bbcb1be0bf3fedd8a6c5cc1a27e4ac07052c239.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/01233d2f05f6d6cfd452a92e97d340101e2cd1993071735db4dc568625dcf9c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "e9b6c82d-7ec9-5026-81f0-dc9c3b1bdcc3", "validation": {"axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004021550994366407, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e9b6c82d-7ec9-5026-81f0-dc9c3b1bdcc3"}}
{"actual_requested_output_tokens": 1813, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 47, "NP": 47, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8e78f3e551f869aaa11ebffa9192c4381220dcd1d7e7a981038ca33294fc8e4d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/69f2a12889bde91b36da17d35dfdb783715632c5a70a4b1ec845fdfbb210a803.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/01233d2f05f6d6cfd452a92e97d340101e2cd1993071735db4dc568625dcf9c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/eb18855256de08c885d8c9f6ada718c80dde45a116f5ea9930864ee086d957fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "01a76a35-bfc3-575f-be02-c7ceb7345806", "validation": {"axes": {"B": 1, "MP": 47, "NP": 47, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001481447834521532, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "01a76a35-bfc3-575f-be02-c7ceb7345806"}}
{"actual_requested_output_tokens": 1813, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/00457701c2760b07597c721a6cf8eca76e09885567b38ce4c6077317116f7d18.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/01233d2f05f6d6cfd452a92e97d340101e2cd1993071735db4dc568625dcf9c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "565295c5-3212-59ff-a70b-de6a10f2bf0e", "validation": {"axes": {"B": 1, "MP": 47, "NP": 47, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0041375732980668545, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "565295c5-3212-59ff-a70b-de6a10f2bf0e"}}
{"actual_requested_output_tokens": 1813, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 61, "NP": 61, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9fe6f6a9ea18ba257c7530877a8eb9ce9cb3f597b5166d1e1e2438a1a496161c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3fda55ded9683bf54b58f741ca9a30933924d1d04a464dbe9bb660d3f80533a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/273aad7a6eb9604197b7e0d353b31d5d1893e0afd7c286b60bd6c63500bc57da.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/780d764a36f53117017a46399caba1e5ffbd0241cf65eebfffae85513231113d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3850], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 905, "uuid": "d5d5357b-a650-5d96-8d49-a12fd91e350e", "validation": {"axes": {"B": 1, "MP": 61, "NP": 61, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001480674371123314, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d5d5357b-a650-5d96-8d49-a12fd91e350e"}}
{"actual_requested_output_tokens": 1813, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 61, "NP": 61, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7e5f224305ce20a8c192ce676733ee51c5a4572e33ab2ed9cf361dc33b3a75f4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/273aad7a6eb9604197b7e0d353b31d5d1893e0afd7c286b60bd6c63500bc57da.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3850], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 905, "uuid": "b26bdb1a-4018-5648-a8d0-f7448a6e00fb", "validation": {"axes": {"B": 1, "MP": 61, "NP": 61, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.003980942070484161, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b26bdb1a-4018-5648-a8d0-f7448a6e00fb"}}
{"actual_requested_output_tokens": 1813, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 75, "NP": 75, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ab844f09bf2d43324028d680fea3b733f07d80606d03952d7fb1e0aaf924d986.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/14feddb780fe261177f1414679495cfcd232317c0f703a5f1aac34db44a45afc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c194c78471835ef4264ed79d15e4d6044b968f51714880eb8983408435f59557.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/236171ade9bb885af664a6118418639c23eca59435c0d0b59b55945dfaa7f73c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4756], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1811, "uuid": "d1c1e46f-2af7-5e2a-b6c0-1aa868d7f510", "validation": {"axes": {"B": 1, "MP": 75, "NP": 75, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016657300293445587, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d1c1e46f-2af7-5e2a-b6c0-1aa868d7f510"}}
{"actual_requested_output_tokens": 1813, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 75, "NP": 75, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6c5eb24cf930cc3abfb8089580ca936b5ee009e126424faf647749a5e72d3b91.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "d9d7f2fbf2a990e13a7eb0202951ed1fe3abf98b5d50c5486aa5db8638e7d7a4", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c194c78471835ef4264ed79d15e4d6044b968f51714880eb8983408435f59557.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4756], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1811, "uuid": "384e9333-58eb-56a0-96c5-5e99088ad842", "validation": {"axes": {"B": 1, "MP": 75, "NP": 75, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.00413424102589488, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "384e9333-58eb-56a0-96c5-5e99088ad842"}}
{"actual_requested_output_tokens": 541, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 232, "NP": 232, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/662f645e3496f813ab8b9ce48c2750d34c7252cb0215e365942cbd4ead13ed76.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cc477f17a90f6040c40a30aa1ead50a15f2bb1074d35206ce27cff3e173b5a31.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0d4d695fff79ad7429f2e59fc1bb8e6aeffd7a3cbaee1765b89b157c4a1104e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/aefad72b625b6b648373587d21f3058a29aa49e31be3a59ce1bd282908db34a6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14849], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "93d0239f-9b63-5ae2-b3dc-c26073264bc9", "validation": {"axes": {"B": 1, "MP": 232, "NP": 232, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003119146917015314, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "93d0239f-9b63-5ae2-b3dc-c26073264bc9"}}
{"actual_requested_output_tokens": 541, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ca5461d09d069bbc79f8d18b02ae873ce3116c019960600a36439bc843f55a91.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2a56c2ac01d0b040ff7c035d4ebeab5ff76f2969434150aba99bd665e02fd546.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14849], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "0909ae91-5e05-5b46-89f8-9a5f7a619db4", "validation": {"axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.00812442321330309, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0909ae91-5e05-5b46-89f8-9a5f7a619db4"}}
{"actual_requested_output_tokens": 541, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 232, "NP": 232, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c557035c561b0008fc7dbb0a0c29fbd9b260a5d5b0c055da53f939590dc0ef9c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bb82c59e0375df14ed94b45be97dfd88dedd683adecadc28ee67bc73a82d6c10.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0d4d695fff79ad7429f2e59fc1bb8e6aeffd7a3cbaee1765b89b157c4a1104e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d58c667fabc51504a8d8b6578e8de73dcdc958106291be2f4036dc872b8e3772.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14850], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "a7fab195-0bdd-55df-92b5-4790bd2720f4", "validation": {"axes": {"B": 1, "MP": 232, "NP": 232, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016978508792817593, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a7fab195-0bdd-55df-92b5-4790bd2720f4"}}
{"actual_requested_output_tokens": 541, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/159a7b0a723952fff649549899722da672030fa57dd8dc89d7560512d2b37211.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2a56c2ac01d0b040ff7c035d4ebeab5ff76f2969434150aba99bd665e02fd546.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14850], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "51038a52-881c-5f07-8469-07ee370e87a9", "validation": {"axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.008127531036734581, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "51038a52-881c-5f07-8469-07ee370e87a9"}}
{"actual_requested_output_tokens": 541, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 232, "NP": 232, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4881a1a01656f079d2a0d46e7e86b1c16fd1984bb4b023d09ee86c6c399a436d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3419fdae2acf6cd7943541b57cd71eca2abd7fce948e2272ebe8dcd68bbacc4b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0d4d695fff79ad7429f2e59fc1bb8e6aeffd7a3cbaee1765b89b157c4a1104e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/50b5298dbf9673c5e749645f7423652345b5d36fe25456178bc210b9922945a7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14851], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "0180c65e-193b-577d-b7a0-0f89f8315c03", "validation": {"axes": {"B": 1, "MP": 232, "NP": 232, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016033519059419632, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0180c65e-193b-577d-b7a0-0f89f8315c03"}}
{"actual_requested_output_tokens": 541, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8b03668561c8c502d70aeadea41c8fa90a590703de83a18913c8bd0f7a9bb037.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2a56c2ac01d0b040ff7c035d4ebeab5ff76f2969434150aba99bd665e02fd546.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14851], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "35a71086-159e-5b06-beb5-fa7d8c1b15ed", "validation": {"axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.008135301992297173, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "35a71086-159e-5b06-beb5-fa7d8c1b15ed"}}
{"actual_requested_output_tokens": 541, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 233, "NP": 233, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0415eeef35746f84fc3e1cc98ec947c91716ddc99515c9412b7c1a90cdb855ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8012ccea4ffb4f49ff2322ed278f7d04ae3a43b438f011dc70e1686a16f08e4c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2a56c2ac01d0b040ff7c035d4ebeab5ff76f2969434150aba99bd665e02fd546.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d265df3954f74da7bed836e4a543f0af463bb21007ba3fd0047972fb6699180.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14852], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "177d0707-dacb-5bfd-b58c-50ee57c71815", "validation": {"axes": {"B": 1, "MP": 233, "NP": 233, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016596848145127296, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "177d0707-dacb-5bfd-b58c-50ee57c71815"}}
{"actual_requested_output_tokens": 541, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/699eea9d794a2aeec9419a8c5db21d178ac0f0bd974ca27f757289780e54a8c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2a56c2ac01d0b040ff7c035d4ebeab5ff76f2969434150aba99bd665e02fd546.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [14852], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "cf71287e-a44b-5bad-b9ee-4ecf1c55708f", "validation": {"axes": {"B": 1, "MP": 233, "NP": 233, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.00811237609013915, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cf71287e-a44b-5bad-b9ee-4ecf1c55708f"}}
{"actual_requested_output_tokens": 541, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 237, "NP": 237, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f35805e8de0d86c3361852ecf1aa8bcadbf2753d8410898b0e1de318a48041cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5bd9692713806b2581e585dcda4c49bf75caa4338006f3fd0358efa2c2a5ae85.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4103bc87b0925b27eb51f7258808e87bbc42d733e60674aa8d51cee156019bd1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ba25f57d37c199e5cb982c1bfa8bb6e000a1f4515fb54f552e3bad7fd8859c4c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [15118], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 269, "uuid": "3d5f6823-d3a2-5ac2-95c4-941958dd16ae", "validation": {"axes": {"B": 1, "MP": 237, "NP": 237, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016633933410048485, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3d5f6823-d3a2-5ac2-95c4-941958dd16ae"}}
{"actual_requested_output_tokens": 541, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 237, "NP": 237, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fed86e925d5cd72b74477352228a4133366930d1b1a2971a8351396530f0e50b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4103bc87b0925b27eb51f7258808e87bbc42d733e60674aa8d51cee156019bd1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [15118], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 269, "uuid": "f1dfc3f3-a819-58a6-a556-cf4ef52ebce9", "validation": {"axes": {"B": 1, "MP": 237, "NP": 237, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.00819517020136118, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f1dfc3f3-a819-58a6-a556-cf4ef52ebce9"}}
{"actual_requested_output_tokens": 541, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 241, "NP": 241, "P": 16}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/292d38a1b7fb914abdf1d133814e34eb16d2f97b8b36a311c204b81b6b057361.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1963099ca4b44757d292b1281b5327ab3de476ac1994f578544c6e201ac74757.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cd2ebba59d0135a69d558ba49bd73bc1d84af0c45053ac0e580cd09470bf488f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b8701ae754735ec884c4f872b2511d8ae5dc97b1b023d6ff0083aaf936946967.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [15388], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 539, "uuid": "5547f7a6-2705-57f9-a9f1-9f6c86ff1f10", "validation": {"axes": {"B": 1, "MP": 241, "NP": 241, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001653146930038929, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5547f7a6-2705-57f9-a9f1-9f6c86ff1f10"}}
{"actual_requested_output_tokens": 541, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_006_268979de", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 14848, "label": "256k_context_p10", "locator": "requests[8].requests[26]", "model": "claude-opus-4-8", "output_tokens": 541, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 10, "target": 14848}, "session_id": "3cb1c6dc71a21714587287306ebe52748568", "t": 648.362, "uncached_tokens": 14848, "uncached_tokens_256k": 14848}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 241, "NP": 241, "P": 64}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 269, 539], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/19a476e49a17eaf08a299272df257594d2bfc659a39a885a27ce4eb3aa2282b1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "fe7b77525307839e408d45ea1d21366dd320b11e4773edc401fa8226a2c615dc", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/cd2ebba59d0135a69d558ba49bd73bc1d84af0c45053ac0e580cd09470bf488f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p10", "scheduler_mode": "decode", "scheduler_sequence_lengths": [15388], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 539, "uuid": "857c26c5-49d9-58a1-824a-5273a6fdcf7c", "validation": {"axes": {"B": 1, "MP": 241, "NP": 241, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0081117725931108, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "857c26c5-49d9-58a1-824a-5273a6fdcf7c"}}
{"actual_requested_output_tokens": 220, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1387, "NP": 1387, "P": 16}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/396e3c47b03b7842fc24becdf204943e718716b630f7a1be697fa16c25ac5eae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1ab4178248aee1dbc0a5a979024d1b552e22defdd247188fa1b0471665f1332e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b73da8c33dac0445b264dc5eddd2a1684082f124bcc01c039bfc05d7b67a272.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7da56d5b1e9e65bba775ef8df0d770840433c0ec9ed3c3ecf4d0b30ed7b4afd0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88769], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "51b0efa4-d7c8-59f5-a9d5-b8ce0cc3463b", "validation": {"axes": {"B": 1, "MP": 1387, "NP": 1387, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.011185440234839916, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "51b0efa4-d7c8-59f5-a9d5-b8ce0cc3463b"}}
{"actual_requested_output_tokens": 220, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ac51c7ce98dbfd32b63e5213825f81dd7146f7b05ab2657a6475304a19b371c8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/32fd6a1ce7b3f5a78d5e9fa79e02c0a07a914ce538933bb8f037e0e6e7620abb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88769], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "f1af6d78-d5d8-5dee-8f36-512ed6a80e0b", "validation": {"axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.10570643004029989, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f1af6d78-d5d8-5dee-8f36-512ed6a80e0b"}}
{"actual_requested_output_tokens": 220, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1387, "NP": 1387, "P": 16}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8aae2fb4154c68bd35a2d8fbbd1e857b4eaba9205ecb0424df65f6f3840b283a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e8bb2917e315ddebb563db22ac24f495333151126e1bd04318cba75474e629ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b73da8c33dac0445b264dc5eddd2a1684082f124bcc01c039bfc05d7b67a272.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/84a01f74bbc70b75cdbc4d245df84d077b5b58ec420228be7d2cef99b04e351b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88770], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "63514b9f-6f5a-5147-bb7b-43097d8ca6db", "validation": {"axes": {"B": 1, "MP": 1387, "NP": 1387, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0026289201341569424, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "63514b9f-6f5a-5147-bb7b-43097d8ca6db"}}
{"actual_requested_output_tokens": 220, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d2b2fed00ea2d26c9e3e1bae569bc40dd8de6eb4e3d6972830dcc6490e2e026f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/32fd6a1ce7b3f5a78d5e9fa79e02c0a07a914ce538933bb8f037e0e6e7620abb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88770], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "a4d18e7b-83e4-526e-87ed-752431476405", "validation": {"axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.03575218887999654, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a4d18e7b-83e4-526e-87ed-752431476405"}}
{"actual_requested_output_tokens": 220, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1387, "NP": 1387, "P": 16}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8cdf1dc77b65ce8e3d46ba58e601c1e48a8e955cdbe3f7ed35ebdf780f4df5e7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f0dddc1a0a5d623034c2555d5ce3a49b912869661d16c175404fba646422481a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b73da8c33dac0445b264dc5eddd2a1684082f124bcc01c039bfc05d7b67a272.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9d8fda9816c6c5a2617cddb2ead698ce9351e80161db7845d7dd72269e5543a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88771], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "d130e370-d1c2-54ea-94f3-9ccc5d5a1fe3", "validation": {"axes": {"B": 1, "MP": 1387, "NP": 1387, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002267216332256794, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d130e370-d1c2-54ea-94f3-9ccc5d5a1fe3"}}
{"actual_requested_output_tokens": 220, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c4c0785770cdecf05729669b516d997c109fdaa222081cc17e6b8874d5f81c20.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/32fd6a1ce7b3f5a78d5e9fa79e02c0a07a914ce538933bb8f037e0e6e7620abb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88771], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "7364ec27-92ef-5282-b794-f110c6d84a5c", "validation": {"axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.03573250211775303, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7364ec27-92ef-5282-b794-f110c6d84a5c"}}
{"actual_requested_output_tokens": 220, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 16}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e295d796b1c4f7174c117b0c90f269dc06dc13921082582c885d1ed16f7ac59c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/514234625d10c7ed2c904ec1030149f45b032809d2bdca61b2d4b56dfa772797.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/32fd6a1ce7b3f5a78d5e9fa79e02c0a07a914ce538933bb8f037e0e6e7620abb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7a00b1c5cef1e05f6009bd8746c16ee1d3707087c16580c42d0f2b2f3113438b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88772], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "36f1a333-2410-5cec-9544-32f19d75d0ee", "validation": {"axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0026582409627735615, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "36f1a333-2410-5cec-9544-32f19d75d0ee"}}
{"actual_requested_output_tokens": 220, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/54f6c5d20477639de3da074c5206d805cf53c604f68c32ed82727e62ed206a93.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/32fd6a1ce7b3f5a78d5e9fa79e02c0a07a914ce538933bb8f037e0e6e7620abb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88772], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "d51d9cf5-d0c4-595c-9d4e-fe380257c72d", "validation": {"axes": {"B": 1, "MP": 1388, "NP": 1388, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.036146556958556175, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d51d9cf5-d0c4-595c-9d4e-fe380257c72d"}}
{"actual_requested_output_tokens": 220, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1389, "NP": 1389, "P": 16}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/75bec3eba84cf1877f2f2ad4482be91c62393530696b7993fc3b0d1823b362c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d4a387f698cbf850d45d8e201bcf6afb24e1ca660cca9a3e1fff450708321ce5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/283679b81951663d622c383dfde6d094df2a18c410e290d442dd77783e213b02.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0d372fef1797736fd0437ac0b9ddfd04dc56543f10084b65cf71eab7219a4e26.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88878], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "ace51816-70ab-513f-9450-0fcf9ff7a24f", "validation": {"axes": {"B": 1, "MP": 1389, "NP": 1389, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002510346006602049, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ace51816-70ab-513f-9450-0fcf9ff7a24f"}}
{"actual_requested_output_tokens": 220, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1389, "NP": 1389, "P": 64}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6a5002b73bbc4f52e919125457ea800a0a81d4bbadf4630d34ad234ad388085e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/283679b81951663d622c383dfde6d094df2a18c410e290d442dd77783e213b02.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88878], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "a84060ab-e1f3-5fd2-858b-b28bccda9f98", "validation": {"axes": {"B": 1, "MP": 1389, "NP": 1389, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.03371073678135872, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a84060ab-e1f3-5fd2-858b-b28bccda9f98"}}
{"actual_requested_output_tokens": 220, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1391, "NP": 1391, "P": 16}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/08ab8d4c9518cddd8a253b08a430bfd4b4c10921ed1331b1fccd1e96561e801e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b65bbd629da49fd8e7cbdf1a02db4e097117f7f3ccf231eb957014430d11594f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9037ffc30d684c4e6e3202e5a86342be69322242e889fb730c9135a5fd3b8ca0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/45ba04a83ce4c5a84361c7082521b83176be688b2967c6be40d7864dc72c63b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88987], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 218, "uuid": "33a51133-92b2-5b13-b7e1-f619f92df9d0", "validation": {"axes": {"B": 1, "MP": 1391, "NP": 1391, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0025443728081882, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "33a51133-92b2-5b13-b7e1-f619f92df9d0"}}
{"actual_requested_output_tokens": 220, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_011_3e09e1e2", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 88768, "label": "256k_context_p50", "locator": "requests[132].requests[6]", "model": "claude-opus-4-8", "output_tokens": 220, "reusable_prefix_tokens": 88064, "reusable_prefix_tokens_256k": 88064, "role": "subagent", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 88768}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 43280.115, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1391, "NP": 1391, "P": 64}, "cache_predecessor_locator": "requests[132].requests[4]", "capture_decode_steps": [0, 1, 2, 3, 109, 218], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0d59f604988f48a1f059d848d98c11fb8f8bfd045d9b04fbb31d2cc3a8ec8ba9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "25379ae90e70cc3c38bddb49f7eddd613857b31c844327371cd51a89e97a1518", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9037ffc30d684c4e6e3202e5a86342be69322242e889fb730c9135a5fd3b8ca0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [88987], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 218, "uuid": "36efc1f0-8760-533e-b4c2-09993171a242", "validation": {"axes": {"B": 1, "MP": 1391, "NP": 1391, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03192408615723252, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "36efc1f0-8760-533e-b4c2-09993171a242"}}
{"actual_requested_output_tokens": 4650, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3192, "NP": 3192, "P": 16}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9cbb6fdf05a98f13c6d8625f0f24d37164e400533476f4221267e147e862abab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a67d71ada0f454630f436e2a1251ee300985e0f65d52787d06d6247b35c0c0a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6d0f89a2afe0e331746c8ce35625dc0e4d694ff97c424cbc42181b372485569d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8e1dab367d3c555c8f1cc0e6371af480d21de419cbd7314bf7dfedd10d22c40a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204289], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "28030388-b731-5105-ad66-6939cd888f51", "validation": {"axes": {"B": 1, "MP": 3192, "NP": 3192, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.009513027034699917, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "28030388-b731-5105-ad66-6939cd888f51"}}
{"actual_requested_output_tokens": 4650, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3e31269d40971365f34ff5f673c546a16106b1b978a3a6249059f3c3aa85c05f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6355d0e55862a045cf095dec9bcef6c5d63fd39f32f2ee98a4ddf18f98bb3f90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204289], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "ddd92acb-4b1d-5b61-9eba-dee6bc202b88", "validation": {"axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.1817234018817544, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ddd92acb-4b1d-5b61-9eba-dee6bc202b88"}}
{"actual_requested_output_tokens": 4650, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3192, "NP": 3192, "P": 16}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/31b6f06bcb47560ffc1d6774382a509a87871c738ee244a817cf8a6534704932.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bfdcb802164a877a054803d6b10e60018175364c09274ec880baab6eb57ea2d5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6d0f89a2afe0e331746c8ce35625dc0e4d694ff97c424cbc42181b372485569d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c5289916c22f7c1bbd264baee9e2045317352553f63427daecc2702c274a0b19.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204290], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "dfb27bea-569e-53d2-af79-6c3def980627", "validation": {"axes": {"B": 1, "MP": 3192, "NP": 3192, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0037358910776674747, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dfb27bea-569e-53d2-af79-6c3def980627"}}
{"actual_requested_output_tokens": 4650, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8a2f053903169f6998ae06990349849c04435a52df65be00b81cebe5b56b5c8b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6355d0e55862a045cf095dec9bcef6c5d63fd39f32f2ee98a4ddf18f98bb3f90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204290], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "70d20fa1-e107-5fa0-8e60-f9adc940d12e", "validation": {"axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.05996692879125476, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "70d20fa1-e107-5fa0-8e60-f9adc940d12e"}}
{"actual_requested_output_tokens": 4650, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3192, "NP": 3192, "P": 16}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/86369e266d144aceccbcd8fedc939061d185e50637e3be5e0926183cdcd78a3b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ebfdd58620c4829fe2b00078c824c134f4c663d06a7c12c170f6beca83515d5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6d0f89a2afe0e331746c8ce35625dc0e4d694ff97c424cbc42181b372485569d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/30436320ff3379beaa94d5a7198005712cbfa4ac371e86ef6d1b49ab68942333.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204291], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "940c8738-ab22-52f9-855c-7ec844c5b9dc", "validation": {"axes": {"B": 1, "MP": 3192, "NP": 3192, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0032376889139413834, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "940c8738-ab22-52f9-855c-7ec844c5b9dc"}}
{"actual_requested_output_tokens": 4650, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c38739533ae443e1c44c5440d604749fb46fb6f1a915e711b01f0790c69dbfc4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6355d0e55862a045cf095dec9bcef6c5d63fd39f32f2ee98a4ddf18f98bb3f90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204291], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "6ec97c4a-7f48-5823-a59c-765e63f13261", "validation": {"axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06002944242209196, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ec97c4a-7f48-5823-a59c-765e63f13261"}}
{"actual_requested_output_tokens": 4650, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 16}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3326fe01f2331cbdef438f61337cb8bcde1971f6dc9b555a0b91c1d9d049af13.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6163d9bcd50e1f70cf4699c5db4a6333fa9967ffd15dd941d7e25d7dd8af4613.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6355d0e55862a045cf095dec9bcef6c5d63fd39f32f2ee98a4ddf18f98bb3f90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/53605af0f06cf0fc9069c7852dc2464277af036b354db663959325a7e8b1f1b7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204292], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "abae23b7-b381-5678-b4e6-685fd7812010", "validation": {"axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003703807946294546, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "abae23b7-b381-5678-b4e6-685fd7812010"}}
{"actual_requested_output_tokens": 4650, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1153281d36111465b91cff06b31a4801524f0ed6dd99400582e6189c38c3e5dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6355d0e55862a045cf095dec9bcef6c5d63fd39f32f2ee98a4ddf18f98bb3f90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204292], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "fcb94950-5298-5eb2-9628-68f6d8aa523b", "validation": {"axes": {"B": 1, "MP": 3193, "NP": 3193, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06012942409142852, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fcb94950-5298-5eb2-9628-68f6d8aa523b"}}
{"actual_requested_output_tokens": 4650, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3229, "NP": 3229, "P": 16}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3f18f8023f907bbc14acc3fe2de47deabfe84e10564737a7c6581e5704be9239.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6ab258d38f0b0b272a7c8fea842a8af320745c752901a6825031c9f8b6d88764.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9aa65b8a044bc534bbfe965767bab19d4d24b7dec15dd14e99a80be9c02c1bd7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4fed60884fb32ff0102d081fc63079b9a79894f5f8276bad6492dd9c09e4fcea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [206613], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2324, "uuid": "8c4db41d-074f-572b-b4fd-fdd6e0ae94b9", "validation": {"axes": {"B": 1, "MP": 3229, "NP": 3229, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004080232232809067, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8c4db41d-074f-572b-b4fd-fdd6e0ae94b9"}}
{"actual_requested_output_tokens": 4650, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3229, "NP": 3229, "P": 64}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5c3ad6f19bd4154d051b978cc1e7c1451cb6c5374139a6af419cc750e26e7cb4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9aa65b8a044bc534bbfe965767bab19d4d24b7dec15dd14e99a80be9c02c1bd7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [206613], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2324, "uuid": "13018c51-e41d-5d89-af55-52dca63bc74a", "validation": {"axes": {"B": 1, "MP": 3229, "NP": 3229, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06261216569691896, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "13018c51-e41d-5d89-af55-52dca63bc74a"}}
{"actual_requested_output_tokens": 4650, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3265, "NP": 3265, "P": 16}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7e84d6069e7795aeadf9c0e5dc7324a1a9ab20154a609a88ebe7cc1dbd6e4181.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c4d63450ef195964b79aa6ae6cd69101585d8d629d7fde298396eef2b7b92210.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8118e20ba94ba8848a6fd5524e44f9b1e719bd96ee03167465bbae4eccb08a8f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e36185518a6d0bc4bf35909442c360d992e28c8a5c20925d5ce73a31457a9933.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [208937], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 4648, "uuid": "28fa3018-be21-5ab2-9961-2a40aa61ac4b", "validation": {"axes": {"B": 1, "MP": 3265, "NP": 3265, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004081800580024719, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "28fa3018-be21-5ab2-9961-2a40aa61ac4b"}}
{"actual_requested_output_tokens": 4650, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204288, "label": "256k_context_p90", "locator": "requests[123]", "model": "claude-fable-5", "output_tokens": 4650, "reusable_prefix_tokens": 201920, "reusable_prefix_tokens_256k": 201920, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 204288}, "session_id": "0b7f3ca692f15ae2651a310d9411d70794aa", "t": 6951.47, "uncached_tokens": 2368, "uncached_tokens_256k": 2368}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3265, "NP": 3265, "P": 64}, "cache_predecessor_locator": "requests[122]", "capture_decode_steps": [0, 1, 2, 3, 2324, 4648], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a8b3fd6d9ea200f2980130af2d48e979ed6635161aec93ba0bd04fd3c4ad1eaa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "92af26998092b30feffc3a59e6f72d495ff6eb76d4464353cd79d3565fc147d0", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8118e20ba94ba8848a6fd5524e44f9b1e719bd96ee03167465bbae4eccb08a8f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [208937], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 4648, "uuid": "95dfea39-b521-5da4-b95c-6542c0eda152", "validation": {"axes": {"B": 1, "MP": 3265, "NP": 3265, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06675602402538061, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "95dfea39-b521-5da4-b95c-6542c0eda152"}}
{"actual_requested_output_tokens": 137, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3892, "NP": 3892, "P": 16}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1bf5f183710c331b30f596087e1fa4ff5acb5c00aaf77437a1ea2e388f667a8b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bb8d77fa73e41cfb21f9b054616714bc434a618f1ceb264e548496a8e91736cb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/72d19a911bd17acf8ed3f2854fe11b322c7985b440640b7a4e7c35d812c16434.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/56c54af0e4dea5f7e8788ce303970faab38aaec4a4538659164f3761019a5c1c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249089], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "e61e222e-fd14-56a0-a233-09c5969aecfa", "validation": {"axes": {"B": 1, "MP": 3892, "NP": 3892, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004623292945325375, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e61e222e-fd14-56a0-a233-09c5969aecfa"}}
{"actual_requested_output_tokens": 137, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c70e54ca5a30aa61877a01f59db5c006fbeac3e60a98c20a30d6b94e799d962b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/70b8da7de47b510530694dbc751778659952fe782f5c1ac6827bac3e580d342d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249089], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "5846e890-6e44-5556-aa7a-96f92f9ffe5d", "validation": {"axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.07924049627035856, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5846e890-6e44-5556-aa7a-96f92f9ffe5d"}}
{"actual_requested_output_tokens": 137, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3892, "NP": 3892, "P": 16}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3b16205d902f40f78d20f1bec6d4aff275cbbbc7220d3d72a1c41355cd1a10ed.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/67cf3f38f6135be96efd7578c8c31c9f054da96b352d7de88b4059a530f010c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/72d19a911bd17acf8ed3f2854fe11b322c7985b440640b7a4e7c35d812c16434.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b3716ff7218e575ef4af2e1951a3b4b540d441f941c46e2c99ce90e83a00b22b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249090], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "6e76e78b-6c06-50c5-996f-b2eff3a5cea0", "validation": {"axes": {"B": 1, "MP": 3892, "NP": 3892, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004550819750875235, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6e76e78b-6c06-50c5-996f-b2eff3a5cea0"}}
{"actual_requested_output_tokens": 137, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5edae94407d8553a48532418d269e81772719aeb61180c7854f7373763256afa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/70b8da7de47b510530694dbc751778659952fe782f5c1ac6827bac3e580d342d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249090], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "24b48023-d883-51e3-8662-feccbf944984", "validation": {"axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0785408727824688, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "24b48023-d883-51e3-8662-feccbf944984"}}
{"actual_requested_output_tokens": 137, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3892, "NP": 3892, "P": 16}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1681b1b00a3b84cfbab404c320b8f313033c357c7a259ac34d85c1a22543f216.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/eef8f6853b9f580742dec168319bf88707e961f955952aa3cfdc69a1a8941a53.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/72d19a911bd17acf8ed3f2854fe11b322c7985b440640b7a4e7c35d812c16434.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f04595c9234545d53a1a419a4e915844d1b34fdf7f84f2c7c90c25f8033959ef.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249091], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "d179ea37-384c-5d0b-849a-e2fe6ec6dcba", "validation": {"axes": {"B": 1, "MP": 3892, "NP": 3892, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.00365437101572752, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d179ea37-384c-5d0b-849a-e2fe6ec6dcba"}}
{"actual_requested_output_tokens": 137, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/54db558ce0d6246d37fb1f282b8a8820bcc6a84a521edb884cd811be771999ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/70b8da7de47b510530694dbc751778659952fe782f5c1ac6827bac3e580d342d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249091], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "1a74cb06-cb9c-59da-8531-e9ba774fcdcf", "validation": {"axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0785565166734159, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a74cb06-cb9c-59da-8531-e9ba774fcdcf"}}
{"actual_requested_output_tokens": 137, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 16}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fdc0ca557ee1966452a246cdaca0dbcecb54b7da580aff8590857d8f7618d1b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ea902f3cfd6e3d078504631dfbf90c052f0eb5b61fc84c49a163117c9681cdf9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/70b8da7de47b510530694dbc751778659952fe782f5c1ac6827bac3e580d342d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c2400cab26bc8e593f37782432c2d8cc50b472396d9c0bc46669f65786e4f300.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249092], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "29b8aa9b-0598-5996-98c4-6ac758422c5b", "validation": {"axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004526659846305847, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "29b8aa9b-0598-5996-98c4-6ac758422c5b"}}
{"actual_requested_output_tokens": 137, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c88f4c2305e1ad156285b8ce30a57bf13098659215f5e8eeb832872eb3ea5646.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/70b8da7de47b510530694dbc751778659952fe782f5c1ac6827bac3e580d342d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249092], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "53777c22-852e-5840-b1fa-f73ed356c31a", "validation": {"axes": {"B": 1, "MP": 3893, "NP": 3893, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.08030253509059548, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "53777c22-852e-5840-b1fa-f73ed356c31a"}}
{"actual_requested_output_tokens": 137, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3894, "NP": 3894, "P": 16}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9fc168c80178a04834aaf66b68d52f578ebf5d7dd9a83acb034770c4f31ff7b7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/60fd6ee732f35968e912ee6c09ee1483d13e4139206e29cc9d062dda9d508bee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/83660356c9b5128afe15a739a1f1a9392eab1757729446314adc5a88bc27d24c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/01a0d457232429e56079f0bc4abfac376be0d0ce5df75233ee503b4e9f4c19b5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249156], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 67, "uuid": "2189730e-bc5e-56c2-a297-42b6d4ed4131", "validation": {"axes": {"B": 1, "MP": 3894, "NP": 3894, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0042535350657999516, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2189730e-bc5e-56c2-a297-42b6d4ed4131"}}
{"actual_requested_output_tokens": 137, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3894, "NP": 3894, "P": 64}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c0ba73aaef7ceeb3ad47fe72a9a0245284b68ac670f88ad268a028187bb4d2f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/83660356c9b5128afe15a739a1f1a9392eab1757729446314adc5a88bc27d24c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249156], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 67, "uuid": "bd886bea-1940-58ed-8b42-0ab8b1730ec9", "validation": {"axes": {"B": 1, "MP": 3894, "NP": 3894, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.08078833390027285, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd886bea-1940-58ed-8b42-0ab8b1730ec9"}}
{"actual_requested_output_tokens": 137, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3895, "NP": 3895, "P": 16}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3c3f9d9e3e4c6af8331e37d4a55a3fa9a5fc66f6efd67e195e6fed2fa67c5ad5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b82a11ee27c1116ca71cbad403eedf6fb953618390b307c72a166f861cada6e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0efc2f9644316e605ce829d27cbc684a8eb5397743cf821788f58402bf16055c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/42b2bac07747c95260a33f19a5408b8a62ee965d8d7a8759ff362170b5e24056.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249224], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 135, "uuid": "55f97d34-37c5-568a-bc45-269b1008bfb7", "validation": {"axes": {"B": 1, "MP": 3895, "NP": 3895, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004829199053347111, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "55f97d34-37c5-568a-bc45-269b1008bfb7"}}
{"actual_requested_output_tokens": 137, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 249088, "label": "256k_context_p99", "locator": "requests[54]", "model": "claude-opus-4-8", "output_tokens": 137, "reusable_prefix_tokens": 248384, "reusable_prefix_tokens_256k": 248384, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 99, "target": 249088}, "session_id": "6ec3296ff5a1f4f0ed4507093863aa53a4cc", "t": 5282.044, "uncached_tokens": 704, "uncached_tokens_256k": 704}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3895, "NP": 3895, "P": 64}, "cache_predecessor_locator": "requests[53]", "capture_decode_steps": [0, 1, 2, 3, 67, 135], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/900704a1803a8ee206a13abc73f75844fa2c5a22d9bdc2946930392983f51ea8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "da5bd76096b64606a3ed18f2450dfe15e7efd728391cb7ee23cad598c3be5a87", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0efc2f9644316e605ce829d27cbc684a8eb5397743cf821788f58402bf16055c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_context_p99", "scheduler_mode": "decode", "scheduler_sequence_lengths": [249224], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 135, "uuid": "6e544100-28fe-517e-8369-a6f4e8b214ae", "validation": {"axes": {"B": 1, "MP": 3895, "NP": 3895, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.08226146223023534, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6e544100-28fe-517e-8369-a6f4e8b214ae"}}
{"actual_requested_output_tokens": 193, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1000, "NP": 1000, "P": 16}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/060a8c872e290313a101de5c13a55b3fd70a2bb3c9a83322b841b2eb73aed611.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8d11eecba1deb6dffa866fcc7ede3a6498ec9b7536857b0e307bedf9908514c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d774fc364c4c0fa164a6c4ad4d7990974595d457015a19db3d32be156af64f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/07293a38421e900ab732b0b0a524df4396d85f565a753dc587faed87a392cf35.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64001], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "c6bc71fe-53e4-5a44-b238-51390bdeefc7", "validation": {"axes": {"B": 1, "MP": 1000, "NP": 1000, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0025354018434882164, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6bc71fe-53e4-5a44-b238-51390bdeefc7"}}
{"actual_requested_output_tokens": 193, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/603446c6012527c7e07c82153a1b3a81febff910b85e8448b776c378a870d1dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e3e7215d4223cdf9f99a074830b84050b73bae827a82d6ea221e9b5cbcbe4bb6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64001], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "3adbcfcd-b781-5856-99db-fa1e0c7b1bb5", "validation": {"axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0313498848117888, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3adbcfcd-b781-5856-99db-fa1e0c7b1bb5"}}
{"actual_requested_output_tokens": 193, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1000, "NP": 1000, "P": 16}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/54c446cd5818e34d4ebf7786e633b28f32f568a5eca73d576ffa49a724ba2aa9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a75c028173c5a9500c1c88e859233e2d719a821afb1c7a9a90d5bd2df7d9b2a2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d774fc364c4c0fa164a6c4ad4d7990974595d457015a19db3d32be156af64f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6064f966a36f30bfc09fb7bb831f168ab80dfccda7327c72bc1839f2c1ee0231.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64002], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "892835a3-ccf5-5459-ba6a-a1c4c1339c07", "validation": {"axes": {"B": 1, "MP": 1000, "NP": 1000, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0024505341425538063, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "892835a3-ccf5-5459-ba6a-a1c4c1339c07"}}
{"actual_requested_output_tokens": 193, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/db2fa5950ee8a5012141a46fa4f487120a70fa04deb81c44e98e9f74852d532a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e3e7215d4223cdf9f99a074830b84050b73bae827a82d6ea221e9b5cbcbe4bb6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64002], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "53eed7c4-3b12-5e88-a376-edf415acd476", "validation": {"axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.028265519067645073, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "53eed7c4-3b12-5e88-a376-edf415acd476"}}
{"actual_requested_output_tokens": 193, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1000, "NP": 1000, "P": 16}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ce491a311618e639ba86d00e8369ee9895c076a5f5a1ecdc6a923655c9416c2e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5f4d5c9cf3ac4504b1ea43308865eac113eb069e7570f0024f079ab6bdfba942.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4d774fc364c4c0fa164a6c4ad4d7990974595d457015a19db3d32be156af64f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/98cce40104c5bc1c732d2a0619a0313ff78c73b336dd6adf59d3975f95646719.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64003], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "e86d7964-cd2d-558c-b83b-2c1e151d3406", "validation": {"axes": {"B": 1, "MP": 1000, "NP": 1000, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002124994993209839, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e86d7964-cd2d-558c-b83b-2c1e151d3406"}}
{"actual_requested_output_tokens": 193, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c237dae5f8cbe0ace9fc13290facb55737eadb9e1bc240754092f5dc48b59615.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e3e7215d4223cdf9f99a074830b84050b73bae827a82d6ea221e9b5cbcbe4bb6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64003], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "47b48c45-fef6-5e5b-b2eb-0a4a69f1c61d", "validation": {"axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.028420817106962204, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "47b48c45-fef6-5e5b-b2eb-0a4a69f1c61d"}}
{"actual_requested_output_tokens": 193, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 16}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/49edf3fd5b215cdc3caee9e4f457a432a1a604009bd6ac1515358354e7b1caac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/67238c1a69e67442c307a0a698fbdcb0d1daaa29e5ba457339b599bc31a8a726.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e3e7215d4223cdf9f99a074830b84050b73bae827a82d6ea221e9b5cbcbe4bb6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/40d8a92671d17925704daf7a0e3eeef808c5ed31395db01df5a22a078a27fcc9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64004], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "d08515ed-4fef-5c03-b171-abaedd5981cc", "validation": {"axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0024454952217638493, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d08515ed-4fef-5c03-b171-abaedd5981cc"}}
{"actual_requested_output_tokens": 193, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e40bc4d2094d93d16c150711ea5553ba09794c59538a0f9a4426e3c931ed6d0b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e3e7215d4223cdf9f99a074830b84050b73bae827a82d6ea221e9b5cbcbe4bb6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64004], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "4f177aa5-2b8d-5476-a548-1eef8ef7e30d", "validation": {"axes": {"B": 1, "MP": 1001, "NP": 1001, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.028274340089410543, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4f177aa5-2b8d-5476-a548-1eef8ef7e30d"}}
{"actual_requested_output_tokens": 193, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1002, "NP": 1002, "P": 16}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a775bfaebf73a9efe6f725de7607ee3541b8f4bd756c14da417b7048fb170db5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9cdc56c15be4d9257198f6939e53897b559816993d2349ea43e2909bf99cbabf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/765392505922356b3c28d529b2b6ac1ae1300ee7f75278d1cec6a64a5795746f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c10c5a0cb606c67dd4fe285abec84a3e12cd2c90d3d1efc7247eb44d92cb4eaa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64096], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 95, "uuid": "e9f6a726-1816-5e29-b18f-774889c859df", "validation": {"axes": {"B": 1, "MP": 1002, "NP": 1002, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0023274091072380543, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e9f6a726-1816-5e29-b18f-774889c859df"}}
{"actual_requested_output_tokens": 193, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1002, "NP": 1002, "P": 64}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2066f709a1bed36d74c65d9352ba17a3d83a125400107d57dc67b7c08ed426a7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/765392505922356b3c28d529b2b6ac1ae1300ee7f75278d1cec6a64a5795746f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64096], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 95, "uuid": "f08cbe2a-8954-5975-b3db-fb7e9b28d819", "validation": {"axes": {"B": 1, "MP": 1002, "NP": 1002, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.02724771900102496, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f08cbe2a-8954-5975-b3db-fb7e9b28d819"}}
{"actual_requested_output_tokens": 193, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1003, "NP": 1003, "P": 16}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4ceed4f125b558677345bfcd18142c3a78140a5fb8e619e04eb3f9350da06232.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e03b5b9e1fe95b1eaa6e887f807c9c9ccad26958ff0ff2335314843a34d903af.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5d90b27cca151dec699ee96131c560acbcf3bbb13ba3fdf8b51ac9d91af16add.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e766e292ca3131567a73670fa58c105c53b18c9b0964dbb86bfe2d47d301cca5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64192], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 191, "uuid": "7a760a1e-8a08-5f5b-9564-c4b10bfc0442", "validation": {"axes": {"B": 1, "MP": 1003, "NP": 1003, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0024389768950641155, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7a760a1e-8a08-5f5b-9564-c4b10bfc0442"}}
{"actual_requested_output_tokens": 193, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": "subagent_014_a5cddccc", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 64000, "label": "subagent_typical", "locator": "requests[142].requests[4]", "model": "claude-opus-4-8", "output_tokens": 193, "reusable_prefix_tokens": 57664, "reusable_prefix_tokens_256k": 57664, "role": "subagent", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 64000}, "session_id": "085aacf44ac5428ad3c6b4f6cd431e083832", "t": 45220.056, "uncached_tokens": 6336, "uncached_tokens_256k": 6336}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1003, "NP": 1003, "P": 64}, "cache_predecessor_locator": "requests[142].requests[3]", "capture_decode_steps": [0, 1, 2, 3, 95, 191], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9b8b440edc361c3443e2242db9fcf8e4ccd07cac49cb07125daa8be8e2ca1454.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "bf5cc8d45196a0b884aa2a8b485cf3e050ecc7ee64af5eaf676b037f22c61712", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5d90b27cca151dec699ee96131c560acbcf3bbb13ba3fdf8b51ac9d91af16add.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_subagent_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [64192], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 191, "uuid": "1b9f5bd2-68e4-5dc4-9395-0470b8578af0", "validation": {"axes": {"B": 1, "MP": 1003, "NP": 1003, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.028276039753109217, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1b9f5bd2-68e4-5dc4-9395-0470b8578af0"}}
{"actual_requested_output_tokens": 4212, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1820, "NP": 1820, "P": 16}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3db09bfc41ceefe04ba90b76e33852046f55165ba3b204912aa280392121d9bc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f8b5f77e587a08e7fc6a54409dd94046f391913f5a4545f4eab471a44a4cfe54.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ab6808990b5d568473adc769aa8c4b5f9bd39955289e107956b0ccb833b15ce7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/83cf37b764f2ed4f76309bda703dcf1f2fdd26ac9f7e137f68d0347989b43b85.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116481], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "3f846c7b-5998-5d4f-a196-10070e080877", "validation": {"axes": {"B": 1, "MP": 1820, "NP": 1820, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003163530956953764, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3f846c7b-5998-5d4f-a196-10070e080877"}}
{"actual_requested_output_tokens": 4212, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3a17b09f242fc771e0d8948f29bc2eec6696cd71f8a6b1896efda2161ce6af1e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d580ce9363dc788ebf4ecaf58186bee95a4038fa6314c661cd4174d500690c7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116481], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "70e8bc4b-637a-5d72-9bfe-1737be10ebf1", "validation": {"axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.05118613597005606, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "70e8bc4b-637a-5d72-9bfe-1737be10ebf1"}}
{"actual_requested_output_tokens": 4212, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1820, "NP": 1820, "P": 16}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6a96492483f3bd2ee0796d2042b5ea04f09ffdd09f0e8972e297b9bec8be064f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b17236442e0e9b053b65c514051b2aff76e8ca953729eb267d3aff3d6271a1d2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ab6808990b5d568473adc769aa8c4b5f9bd39955289e107956b0ccb833b15ce7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/098da355205973484aac34273f90f4336efbeda345af21b4b70fe2354cb75b7d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116482], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "82a57066-d33d-58d9-af4f-e59744830172", "validation": {"axes": {"B": 1, "MP": 1820, "NP": 1820, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0031183678656816483, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "82a57066-d33d-58d9-af4f-e59744830172"}}
{"actual_requested_output_tokens": 4212, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5686c29602b965ceb943f1274787cf86b130f01a8d3cdd2b817def96f3e92083.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d580ce9363dc788ebf4ecaf58186bee95a4038fa6314c661cd4174d500690c7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116482], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "ea744a7d-2ce6-531c-b7ae-77f8bf8203b2", "validation": {"axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04927533119916916, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ea744a7d-2ce6-531c-b7ae-77f8bf8203b2"}}
{"actual_requested_output_tokens": 4212, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1820, "NP": 1820, "P": 16}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f6a70fe3c5ff5712d893ef7c6e6090d975304d17626435df9e855d47e0fa6c7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9ebd98b18c131420e19a8bb194ff5955133e2209436c3d732a24da5e0f063285.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ab6808990b5d568473adc769aa8c4b5f9bd39955289e107956b0ccb833b15ce7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8698ff424853760bcbc42effa866990630d5fc31a2a128acb60630343cf89175.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116483], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "3b57ba87-a331-5823-9478-330eb43ba013", "validation": {"axes": {"B": 1, "MP": 1820, "NP": 1820, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0025400137528777122, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3b57ba87-a331-5823-9478-330eb43ba013"}}
{"actual_requested_output_tokens": 4212, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b47b5d4c7312fd941edc323b374b769c1d576dba7801321fa9fc631272deeeeb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d580ce9363dc788ebf4ecaf58186bee95a4038fa6314c661cd4174d500690c7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116483], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "2b3cac99-c532-55a3-8b38-cdb84077035e", "validation": {"axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04951393371447921, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2b3cac99-c532-55a3-8b38-cdb84077035e"}}
{"actual_requested_output_tokens": 4212, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 16}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/42dc650e4d013742f61eb635af2442a97a0035ffea0305387d2f13fecba73a63.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/857ac5d30f9027d14d4dd57736085723d6e1060b38bd11f3b7e3a0a29d693dcd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d580ce9363dc788ebf4ecaf58186bee95a4038fa6314c661cd4174d500690c7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f001284c55ba53c51273dec9b2bb399ad82082566e82e519e6a9999cebf18df9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116484], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "bcb74036-d932-52f3-a770-169cda01b232", "validation": {"axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003152106888592243, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bcb74036-d932-52f3-a770-169cda01b232"}}
{"actual_requested_output_tokens": 4212, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/090277aaf47188ae71dc09d3944f9d3fce13d2cc5398431f8b91d9ce0a37c66c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d580ce9363dc788ebf4ecaf58186bee95a4038fa6314c661cd4174d500690c7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [116484], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "91971050-cb2f-5e79-93ea-0feace49bc64", "validation": {"axes": {"B": 1, "MP": 1821, "NP": 1821, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.04982542572543025, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "91971050-cb2f-5e79-93ea-0feace49bc64"}}
{"actual_requested_output_tokens": 4212, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1853, "NP": 1853, "P": 16}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/302e3ae8764bbe5a82bb161f5a7eee6053891ee5b21b2b299efdf89bbc23c9a6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e842c820c62bde2169e05db2e5f3cc69d258cb473d73ec5fd78cb51a4f001599.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/974f345b0c679f04c9c7eb05c83716fe4217e91b4c1ddff77c75edd4875d00f4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/54c26b85d2f9f1fe000b890a12e89aa7bff8e0a6a3b6d0d2013c4953bd2250a8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [118586], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2105, "uuid": "e8be0df2-3161-5dff-8de5-b4ada29b03ff", "validation": {"axes": {"B": 1, "MP": 1853, "NP": 1853, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0032056961208581924, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e8be0df2-3161-5dff-8de5-b4ada29b03ff"}}
{"actual_requested_output_tokens": 4212, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1853, "NP": 1853, "P": 64}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b02176524019f08a8fa06919e6c5fd453ece5feb984150122b846c306a6c19d5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/974f345b0c679f04c9c7eb05c83716fe4217e91b4c1ddff77c75edd4875d00f4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [118586], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2105, "uuid": "9fdb533d-b5c3-5b03-bd66-5d32d0ea61b6", "validation": {"axes": {"B": 1, "MP": 1853, "NP": 1853, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.05020872177556157, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9fdb533d-b5c3-5b03-bd66-5d32d0ea61b6"}}
{"actual_requested_output_tokens": 4212, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 1886, "NP": 1886, "P": 16}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/45187eb123b777ca6b3ae7fdbe5df89e9c4373e3a17e785c419e1bf105392d36.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/feae868a366c85b945d8fb65ddf85cf7235ee82881a2078ba674fd42dbf3a0fb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/148536a2c21af9de49b6f580577babb419a684bc4a23153b5216a76e72f5ebac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/797e9fc3d9250c4ae60b501fa78d965dd24203f281206273d74d25858e3c712f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [120691], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 4210, "uuid": "3ffe7dc2-9f9a-58e0-8b01-d87bb66ae20d", "validation": {"axes": {"B": 1, "MP": 1886, "NP": 1886, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0031846663914620876, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3ffe7dc2-9f9a-58e0-8b01-d87bb66ae20d"}}
{"actual_requested_output_tokens": 4212, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 116480, "label": "auxiliary_typical", "locator": "requests[36]", "model": "claude-opus-4-8", "output_tokens": 4212, "reusable_prefix_tokens": 113600, "reusable_prefix_tokens_256k": 113600, "role": "auxiliary", "selection": {"field": "input_tokens", "note": "Role-specific median input length.", "percentile": 50, "target": 116480}, "session_id": "15ab8fdb8831965bb01e45ac51829cb4e3e3", "t": 1598.877, "uncached_tokens": 2880, "uncached_tokens_256k": 2880}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 1886, "NP": 1886, "P": 64}, "cache_predecessor_locator": "requests[34]", "capture_decode_steps": [0, 1, 2, 3, 2105, 4210], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9e4bc8318d9e9942dbe2c6b9c05494fc6997a1d4557ed63277fe6a060e657988.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e0122beba611dd440dcad13c038de367295046efe370303f49935561521bf716", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/148536a2c21af9de49b6f580577babb419a684bc4a23153b5216a76e72f5ebac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_auxiliary_typical", "scheduler_mode": "decode", "scheduler_sequence_lengths": [120691], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 4210, "uuid": "b14735d0-5b69-5ad8-9bf5-75047bc7c273", "validation": {"axes": {"B": 1, "MP": 1886, "NP": 1886, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.05137157905846834, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b14735d0-5b69-5ad8-9bf5-75047bc7c273"}}
{"actual_requested_output_tokens": 204, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3730, "NP": 3730, "P": 16}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6a843b06f26afc84d4ad736fc4176fb4eda950fad53298f71a77ec5d364dd929.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/986c8d3541ecd506472ac4ec8531be4af1966ce9f5b9f1849f1409990385e284.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cb47c2088d07bdbe1eb04dbeb0ef31acde4a6d830b04d93a034391f0de835a6f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/54ecd5290f17c98941c670900c1ca78c21bb0f435729d81fb95a17a1beaa27fa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238721], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "926467dc-f196-57ce-b079-bb8a1d0902f8", "validation": {"axes": {"B": 1, "MP": 3730, "NP": 3730, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004809195175766945, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "926467dc-f196-57ce-b079-bb8a1d0902f8"}}
{"actual_requested_output_tokens": 204, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c5488956eb04e7b07a5fa6706b788357d59f815193480d22dd37348306369d3d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6be55828eeffa5f96268e97cd386b124fffa52eb141bea2854ecae3e7d6b197b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238721], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "e067e72e-25fd-5fc4-80a1-e8fdac27b787", "validation": {"axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07898493018001318, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e067e72e-25fd-5fc4-80a1-e8fdac27b787"}}
{"actual_requested_output_tokens": 204, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3730, "NP": 3730, "P": 16}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1bc6d55b8049f2931560453f3375282740be4aee52a3f20decef2e4f1a52dea4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fd91ceeab489db94d5109085d2fc0c9cd4c191749e3dcaf50ffe2a6a471d5ee1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cb47c2088d07bdbe1eb04dbeb0ef31acde4a6d830b04d93a034391f0de835a6f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7edfd74e9b2ea6af11a4cd8341521734e761b8dc7c003f3c7064d3fece00c7c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238722], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "bb80eef0-3529-57a8-a6b7-ab7891f92705", "validation": {"axes": {"B": 1, "MP": 3730, "NP": 3730, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004537228029221296, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bb80eef0-3529-57a8-a6b7-ab7891f92705"}}
{"actual_requested_output_tokens": 204, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/33c47860c92fbfa981cc7d037aed1249f635ce8a71a276cf21f49420a9b5d83e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6be55828eeffa5f96268e97cd386b124fffa52eb141bea2854ecae3e7d6b197b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238722], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "3ce53636-633f-5ec4-806c-d0fdfd0c1f88", "validation": {"axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07900446327403188, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3ce53636-633f-5ec4-806c-d0fdfd0c1f88"}}
{"actual_requested_output_tokens": 204, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3730, "NP": 3730, "P": 16}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/67a9579a27d2af15e1f61c1188db951dc74d6fb7c22ced538d29db82ce7bc63f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f467f7e58099a96fb3fa9a562a92c9f0e8df8c9386ec3c3cc2ee9c1f5bb7f917.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cb47c2088d07bdbe1eb04dbeb0ef31acde4a6d830b04d93a034391f0de835a6f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/049b7fbd0926c4f1db3e0a87633a9b0ac50a422b6e7c99e138541759b654a9ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238723], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "817a3d84-ace4-55bc-9d09-548c89f4cdb5", "validation": {"axes": {"B": 1, "MP": 3730, "NP": 3730, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0035448060370981693, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "817a3d84-ace4-55bc-9d09-548c89f4cdb5"}}
{"actual_requested_output_tokens": 204, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2a55679017673c1ab5e5a75fb20bf167a75d060c43ec3a235c74210144618802.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6be55828eeffa5f96268e97cd386b124fffa52eb141bea2854ecae3e7d6b197b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238723], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "49a4b757-6bb4-5242-841b-7834abe4e55f", "validation": {"axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.07884885976091027, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "49a4b757-6bb4-5242-841b-7834abe4e55f"}}
{"actual_requested_output_tokens": 204, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 16}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/80c8509fb9320547a2dc34b5192b52e43a766c9bb5b9e49fefb4f87afb16d37c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7b06773fbabd2d244e04b0f9de38712914ddc7d4b0819f2799b7bb21a7df060f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6be55828eeffa5f96268e97cd386b124fffa52eb141bea2854ecae3e7d6b197b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f975544b36b636fa6edde1214ebfbd9a5ae4a1bc8ac510ed34054e2d26fa0667.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238724], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "a3b6c690-11d2-5067-b991-3d50c14ba1be", "validation": {"axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004688025917857885, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a3b6c690-11d2-5067-b991-3d50c14ba1be"}}
{"actual_requested_output_tokens": 204, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7b7fd3ee33c955f12cc883cd96d48eb088ac47c46a2b07ac570902ccce021f0f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6be55828eeffa5f96268e97cd386b124fffa52eb141bea2854ecae3e7d6b197b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238724], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "1f03d47f-86cb-5419-aee9-787887d7a758", "validation": {"axes": {"B": 1, "MP": 3731, "NP": 3731, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.07876059738919139, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1f03d47f-86cb-5419-aee9-787887d7a758"}}
{"actual_requested_output_tokens": 204, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3732, "NP": 3732, "P": 16}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/84e3e6e61381ce80de8503337b5adfef39f89a2bd685e6a3e3ffb32d040d6339.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b319739bd49bf59436b6ec74433a6e3e8adedbf105f99827b57cc1e77966ada8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/148f5c5be9a2fc1ea366cba867cb640dc6a6e4e199cf07911c81a6f5f99a9f81.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2728d8fd1062d847cf90ef239ca37e8e2afef7a4e647d0707b82120d429c5a78.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238822], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 101, "uuid": "c7e597e9-d28d-5401-af12-f8a8e6a57fd5", "validation": {"axes": {"B": 1, "MP": 3732, "NP": 3732, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0046705459244549274, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c7e597e9-d28d-5401-af12-f8a8e6a57fd5"}}
{"actual_requested_output_tokens": 204, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3732, "NP": 3732, "P": 64}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a25da6329da17dbf340492182f9bfe9ec236106485e082afe490e7446ce9a7b1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/148f5c5be9a2fc1ea366cba867cb640dc6a6e4e199cf07911c81a6f5f99a9f81.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238822], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 101, "uuid": "17fecf5c-0485-5401-9d9f-92de639fdbf6", "validation": {"axes": {"B": 1, "MP": 3732, "NP": 3732, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07874611997976899, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "17fecf5c-0485-5401-9d9f-92de639fdbf6"}}
{"actual_requested_output_tokens": 204, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3734, "NP": 3734, "P": 16}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3585d98581fe5c32c76f83c52d195223c7ada3ffd2a74ef9c8097f56cca53c61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b05d13b819a278d2a474835c68b851cbc16bd80fce6f2376fcda1fc679862caa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/403dcf2946d9365ede80a1cd6646135feb34454bc47940172612d1f9bc9704ef.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/22d4008dfacbd418bcc5d4e820ee9223cd679352fe9607f5a36ec66c020d4a8a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238923], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 202, "uuid": "d2ea18ad-71d2-5ddf-8924-c1c747dd4e75", "validation": {"axes": {"B": 1, "MP": 3734, "NP": 3734, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004651776980608702, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d2ea18ad-71d2-5ddf-8924-c1c747dd4e75"}}
{"actual_requested_output_tokens": 204, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 238720, "label": "fully_reused", "locator": "requests[63]", "model": "claude-opus-4-8", "output_tokens": 204, "reusable_prefix_tokens": 238720, "reusable_prefix_tokens_256k": 238720, "role": "main", "selection": {"field": "input_tokens", "note": "Ideal fully reused prefix; decoding only, with no zero-token prefill workload.", "percentile": 50, "target": 238720}, "session_id": "748b75ff61b2e460c3a0da7db15be271afa0", "t": 17383.206, "uncached_tokens": 0, "uncached_tokens_256k": 0}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3734, "NP": 3734, "P": 64}, "cache_predecessor_locator": "requests[62]", "capture_decode_steps": [0, 1, 2, 3, 101, 202], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ad27470bfb6a7149f25072415e5a3f87bc22416fd4ddcb99dacc6335b39960db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "747794bbd6fd6bbbe3a9d21adb1984a3401a9aafc3cec39b5bb01196850664ca", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/403dcf2946d9365ede80a1cd6646135feb34454bc47940172612d1f9bc9704ef.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_fully_reused", "scheduler_mode": "decode", "scheduler_sequence_lengths": [238923], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 202, "uuid": "f9cf81d7-0439-5d1f-a5bb-9b99f373cfa7", "validation": {"axes": {"B": 1, "MP": 3734, "NP": 3734, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07874260982498527, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f9cf81d7-0439-5d1f-a5bb-9b99f373cfa7"}}
{"actual_requested_output_tokens": 7948, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3197, "NP": 3197, "P": 16}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/82223fbf31317ac9206291f084afb52315c3571bb673f398d5d1d086bc7ced0a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8a7579a51f4c700be97f609cc3d9aa016fc10e75fdc03249f30917b6c86b2442.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b4bbc7fe7e10ab57bc8f76fc3a427034e83f31303e39539214239a03add9f48.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c6368deea65ae9913e161b8dc679f462b42725d781f69ec2664e6bb02ea84c59.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204609], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "e61fe529-6b78-5fc4-a702-49ec2a9ebe09", "validation": {"axes": {"B": 1, "MP": 3197, "NP": 3197, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004199176095426083, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e61fe529-6b78-5fc4-a702-49ec2a9ebe09"}}
{"actual_requested_output_tokens": 7948, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/64506deb96f60acd7da7aaf4ce65e9b47d4f964cf6fade25e3d33be775c14e2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a076f28b084c1444858ce2757f751175487900cb073bf362e66892b5bfa41ec5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204609], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7e94392a-ac3a-5de3-8947-99f5081846cb", "validation": {"axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.06801652116701007, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7e94392a-ac3a-5de3-8947-99f5081846cb"}}
{"actual_requested_output_tokens": 7948, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3197, "NP": 3197, "P": 16}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9bb3f6545deb9c52d6dab8d4db71d22707f50de69263de778afbee465257bd6f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2164e09b604435abd858cd18a44d7a7061c4aecadb05ac1ebdc7cfe66a5a40c9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b4bbc7fe7e10ab57bc8f76fc3a427034e83f31303e39539214239a03add9f48.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3ee12ddf80bed23f53664b1bce267876e90b6a6cfac923f26978e3fb943ef5ad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204610], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "69d5118d-b334-53fe-901f-1211d85b5348", "validation": {"axes": {"B": 1, "MP": 3197, "NP": 3197, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004122375976294279, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "69d5118d-b334-53fe-901f-1211d85b5348"}}
{"actual_requested_output_tokens": 7948, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1a2ed487d0c2a75769fe713e7f3cb4140029b94c8b1c6b488ac2b660fe5cd4e9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a076f28b084c1444858ce2757f751175487900cb073bf362e66892b5bfa41ec5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204610], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "6bb78b8b-f361-5255-be07-a9582bbc0e7b", "validation": {"axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06822989229112864, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6bb78b8b-f361-5255-be07-a9582bbc0e7b"}}
{"actual_requested_output_tokens": 7948, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3197, "NP": 3197, "P": 16}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/827b33041b66da579c8f7670ce5ab3564057aafc92b616ae0ecd47b49ab04319.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/307b2c413b1f1c149b66652285c0a4cd96892840e6a021dd2a6a033be0b5912c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b4bbc7fe7e10ab57bc8f76fc3a427034e83f31303e39539214239a03add9f48.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5a87bae060207ba7af73cd472160a77b4f65b97e2d9273dad947c038efaedc64.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204611], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "3d189464-a427-5052-afa8-5f8c81908b2f", "validation": {"axes": {"B": 1, "MP": 3197, "NP": 3197, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0032684030011296272, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3d189464-a427-5052-afa8-5f8c81908b2f"}}
{"actual_requested_output_tokens": 7948, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fa1983d0efc9604d98924c788650ff5c853ba046dc538d257e3efc785e329a60.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a076f28b084c1444858ce2757f751175487900cb073bf362e66892b5bfa41ec5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204611], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "266650a3-3bd7-58ee-9c42-81b2f13c408c", "validation": {"axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06832638382911682, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "266650a3-3bd7-58ee-9c42-81b2f13c408c"}}
{"actual_requested_output_tokens": 7948, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 16}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/11631fce3e3d15afaa3bfee8952a899ed4dd02b202635b7978dde11b5d4cde0d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bef9205bca45a54d75a8b1142b658660f6137cf95ca7c5ea94f259ae509457fa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a076f28b084c1444858ce2757f751175487900cb073bf362e66892b5bfa41ec5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cb9758775e3fe40f7c94325eaf74311aac659678b11893e4992a8f6ad58088ac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204612], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "e6df16f3-922e-5a74-84be-1dddbf86b82e", "validation": {"axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004186600912362337, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e6df16f3-922e-5a74-84be-1dddbf86b82e"}}
{"actual_requested_output_tokens": 7948, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e1d345f4295ff408e02e5cc14feaca5a42b0055929f89ab9aadc7f2545ca484b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a076f28b084c1444858ce2757f751175487900cb073bf362e66892b5bfa41ec5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [204612], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "60e90d02-a2c6-5040-bb1f-60bd2d853ece", "validation": {"axes": {"B": 1, "MP": 3198, "NP": 3198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.06825287826359272, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "60e90d02-a2c6-5040-bb1f-60bd2d853ece"}}
{"actual_requested_output_tokens": 7948, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3260, "NP": 3260, "P": 16}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/06f7bf0affac00f0001229501575ac6a171fc041a8dc6fb66db5db8c7fe82588.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/309535b4df9e5303ad795aeea2d752956a9ebd068869fb26a0a4e4de37cbb73f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/728abf9804b05188f44c0b8973f76bbbce81a95c01149d641567bf0adced679e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cd00fbd8eb359ee486f4c26bdf447721c6e0c1e58ef871f88a716e666948820d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [208582], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3973, "uuid": "63e480e0-3df1-5fc7-9998-4d2c23d114dc", "validation": {"axes": {"B": 1, "MP": 3260, "NP": 3260, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004325061105191708, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "63e480e0-3df1-5fc7-9998-4d2c23d114dc"}}
{"actual_requested_output_tokens": 7948, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3260, "NP": 3260, "P": 64}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fe9ec2d951399c86bf0eeaefcf725fb5a0fce6eb1067e21b7a81d910b94adcbc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/728abf9804b05188f44c0b8973f76bbbce81a95c01149d641567bf0adced679e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [208582], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3973, "uuid": "e899448f-1de2-5c76-8e72-63cb75d32a5a", "validation": {"axes": {"B": 1, "MP": 3260, "NP": 3260, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.07243998814374208, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e899448f-1de2-5c76-8e72-63cb75d32a5a"}}
{"actual_requested_output_tokens": 7948, "adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 3322, "NP": 3322, "P": 16}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/54f2bdffa34c41890a61a6d8217ab3c652e9bb81ffde4e6b497b014d3f172ce1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/524150f85d7fc815b296bba4da027f8782a5347ff4a1702cac1503c22ac3913d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2d9f6c8e79d56cc512b2d484a6eda8bde1b6b7c4f80a4bbb93645f60952d46e0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9f87d7a06e613f7a542dce2c8653d115dd932c1d8d6e361393bfa97734537011.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [212555], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 7946, "uuid": "1a06c89e-7764-5a29-8664-4f1ee515e676", "validation": {"axes": {"B": 1, "MP": 3322, "NP": 3322, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0043017249554395676, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a06c89e-7764-5a29-8664-4f1ee515e676"}}
{"actual_requested_output_tokens": 7948, "adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 204608, "label": "256k_long_output", "locator": "requests[74]", "model": "claude-fable-5", "output_tokens": 7948, "reusable_prefix_tokens": 204224, "reusable_prefix_tokens_256k": 204224, "role": "main", "selection": {"field": "output_tokens", "note": "Long decode trajectory within the Qwen model context limit.", "percentile": 99, "target": 7948}, "session_id": "1b577bfa02cd2c421c0b59c5cac702b3c287", "t": 3693.72, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 3322, "NP": 3322, "P": 64}, "cache_predecessor_locator": "requests[72]", "capture_decode_steps": [0, 1, 2, 3, 3973, 7946], "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5a28c0b8b8a79b1bcead6f9ea2644a7eed213f5110ffcaffc855a8185c3ad25a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "93a0fe2aa6d6a9252367be8da47bf2577fed128d48c71bd4d3371f3c8f22aa25", "layer_id": 3, "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2d9f6c8e79d56cc512b2d484a6eda8bde1b6b7c4f80a4bbb93645f60952d46e0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_256k_long_output", "scheduler_mode": "decode", "scheduler_sequence_lengths": [212555], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 7946, "uuid": "8827697e-7478-5d3e-b017-ca652334a3b6", "validation": {"axes": {"B": 1, "MP": 3322, "NP": 3322, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07027529133483768, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8827697e-7478-5d3e-b017-ca652334a3b6"}}
{"actual_requested_output_tokens": 107, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 128, "Nkv": 128}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/4b208d445b76a9e22b7ba8aeb15462cabf8a2fb8fea232fa6b891071e1bd855b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [128], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "63072e3a-2bf9-5eed-bc07-c48f323b15d4", "validation": {"axes": {"M": 128, "Nkv": 128}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.014219233300536871, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "63072e3a-2bf9-5eed-bc07-c48f323b15d4"}, "variant": "rows_128_compression_0_cached_False"}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 128}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/cab42d02ef8e1bdab85d77ae269c85693eb1812e200373e99b3924665d385f2f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/c266a1c92d4f6cd3062553bb9c99ff9db32cdade2ad7c4be4424a6614286b082.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a9ef4b85ed35d2174174084a725cccf38c9ef6407f8b937b32ba67ba5cb57acf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7531a7c409a876f61ac259c161074206c934d45aedb50b4ea309d8ea1a80e55b.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [128], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7414061c-d565-5448-9e7a-78d5b1ebfdb4", "validation": {"axes": {"M": 128}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.3113021850585938e-06], "seconds": 4.782656603958458, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7414061c-d565-5448-9e7a-78d5b1ebfdb4"}, "variant": "prefill_128"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 768}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/274614eec2f892ebe8291f84a02c8c52442cf7fc51a0bc0b7244c82b4193f1d4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [128], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "edde6f2e-b9cf-5ae4-b99c-4e313bf6199f", "validation": {"axes": {"M": 768}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 3.690998049918562, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "edde6f2e-b9cf-5ae4-b99c-4e313bf6199f"}, "variant": "prefill_128_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 768}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/6124ed8165865b9abde8ddb3856758f856b53c2b64afc98302b5a615d0e451e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [128], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "59c73c1c-f9e6-54f7-b15f-f6384e355503", "validation": {"axes": {"M": 768}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.7128978883847594, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "59c73c1c-f9e6-54f7-b15f-f6384e355503"}, "variant": "prefill_128_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 128, "Nkv": 192}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/5951560539c31667434e9faf5c9f0c9e7f5e7cbc32296c43ab32eef581c10d3a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [128], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "d9165058-a73b-5e77-8e61-5f92a8458b70", "validation": {"axes": {"M": 128, "Nkv": 192}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.011357225012034178, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d9165058-a73b-5e77-8e61-5f92a8458b70"}, "variant": "rows_128_compression_2_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 128, "Nkv": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/ff38c3cc6092e80e32b3e1ec3adb2fde8aab515b6241d22dcb7752f9fbe76a50.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [128], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "546bcf0e-32bd-5dd8-9ea9-14c58e6a52c9", "validation": {"axes": {"M": 128, "Nkv": 256}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010068903211504221, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "546bcf0e-32bd-5dd8-9ea9-14c58e6a52c9"}, "variant": "rows_128_compression_1_cached_False"}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f18622b409563c96228e8293a5140a6bad27db5cd1ec4521702671c4948fbcbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/875157d7a4b63732f005fccc03c4c154e9e1b4760317fe3c658e8fee3014dc4c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/101a87b53a0f82bc273d8c226c0d729c276ecd4faa4fde842f8cd968f676402c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/41f7eb2d843f705dfd20f53ad4d3220934c0020a5dd2852f221cf0d63daae43e.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [129], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "712fdc6e-f456-507f-9fa3-dbc410c44cce", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 4.470348358154297e-08], "seconds": 0.00314719183370471, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "712fdc6e-f456-507f-9fa3-dbc410c44cce"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/3613db7e6c03bf5b77e75bf0f05a67f5bf71a1ef12e59c0d65e05a1c695a25ef.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [129], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "70fc1ff9-26db-5fb7-8873-070ce7537f56", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5481769633479416, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "70fc1ff9-26db-5fb7-8873-070ce7537f56"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/a9ff30c9fe5aad7f4f45297055b4c7df9736e9ed160969acbde0e74f589187d4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [129], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "b5374401-90e6-5268-96c8-b315c811aa64", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5974064632318914, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b5374401-90e6-5268-96c8-b315c811aa64"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/639c79b0fec07e42b02799a5e8c2136dbfab449686e3b5c21c734566f0c2f23f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [129], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "a141972f-4cf7-5111-967c-2dcb6463b86c", "validation": {"axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.07541759125888348, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a141972f-4cf7-5111-967c-2dcb6463b86c"}}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/541ac6d7956b18859b40d4f8393df652ef15b6d596cd2de588fa1c9a291fbab9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a550e7a4da728acbf6496ca2f1dcd7e77204e926b6109ae241c6c7281db19d97.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/54d2b115bf542bcf4433f9eb38481ee78b79522851f57ed8839d086ce54c1cf4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f3db39033001d5aa6ded5a26ceca2bb4261d8f0b52ef8de213adfd7d1370e373.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [130], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "5d9a5db7-f338-578b-b8b0-8d3d1e5a06c8", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.002588692121207714, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5d9a5db7-f338-578b-b8b0-8d3d1e5a06c8"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/ae27b9f4145d74b3bf24000f1cace3128d66139de78b6b512460d616d11012b0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [130], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "d1e43d5c-b780-543d-bd90-cf2550354d5d", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5461130570620298, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d1e43d5c-b780-543d-bd90-cf2550354d5d"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/72d7d3d27fbf96252492fd0d82886fbe79bcd66fdd5ca5e40aed69435c0e37cb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [130], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "8d1a3998-f38b-512d-aaf3-4d76b87a42f0", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5964415301568806, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8d1a3998-f38b-512d-aaf3-4d76b87a42f0"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/a8c15c2892df53a2ab99563f6897d7d423bdec63e23a36d65f2426bc50c872d0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [130], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "eafd3828-d238-5d93-ad23-6443cd632c0a", "validation": {"axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003597497008740902, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "eafd3828-d238-5d93-ad23-6443cd632c0a"}}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/451565d54bd2a077dfefd6266d8eebf60da524a1aab2bd30316b03a2124c3c72.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/06b06c7ef0de7911f410a31395090d18c9385704026e95428b51d3a33f55468b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/eff5114fd7e82fed274487ad2961094ad2b29e6e1605b17825fe01992842a1b5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/930cd3c1337652be54b78d92e8b4529f05de93451f2c67451a1683b52e76d70a.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [131], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "ea1d4f58-48fc-51af-8df3-e239f1fbddf8", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.001954874023795128, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ea1d4f58-48fc-51af-8df3-e239f1fbddf8"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/b725efa8c9b438dfa626b9811fae2be2dd2cf2be8686c55b1d2c2a3eec425346.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [131], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "f11431ca-994c-55dc-a835-d8cc75635ba8", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5454176180064678, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f11431ca-994c-55dc-a835-d8cc75635ba8"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/adb06105a0fc4e5640b89a12d54d3dff52967c9b9cd39d0b9d30b78c28cab534.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [131], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "658b6b66-5595-55c4-9ca4-fb508696cfbd", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.4939532349817455, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "658b6b66-5595-55c4-9ca4-fb508696cfbd"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/ee0c3e5977b92eb9861bf4f0823a3eaf89448822a51a43cec93c61f6962a5765.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [131], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "3aa1fefe-5390-5840-a649-b5e50c2ffd2f", "validation": {"axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003529265057295561, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3aa1fefe-5390-5840-a649-b5e50c2ffd2f"}}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/626a5d5850cc927b7b73f10a1efca1cadf7438b266d24d4482cc15fc547c5c73.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/e0f79bac5b2e2fdcd3f62e3f6b9dbb5487c660c376405b786a15f5c7a6240533.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/fcae52528ea4033fdb8d8944bb39c26a337caf7f332e1d99b28c61ee0454fec3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/74a7a774ef2576f0fe5dc5efdc0a7ec213cc6a1712997493f5b4d412776dd47a.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [132], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "3e2153a3-11cb-5f0a-bedf-28f7e1b35d27", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0019683837890625, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3e2153a3-11cb-5f0a-bedf-28f7e1b35d27"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/975cca1048447dcffb531bd0614cb335662954a0c8c406f7dc0c8e22146b8468.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [132], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "23dbe7a6-7be5-5b9b-bcdf-8ce53441d95e", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5454356167465448, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "23dbe7a6-7be5-5b9b-bcdf-8ce53441d95e"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/773e8694bc4562765e8352be7d14779de9d03c79fbff751c47bca0ecf25ad732.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [132], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "9502d66f-df7e-5068-96cc-e538986fec52", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.346023241057992, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9502d66f-df7e-5068-96cc-e538986fec52"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/1508164425014a88077326c657593b16b60054b519619b3dc450045b09fa270a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [132], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "3d3493ba-7f57-54e3-94c5-34e37f52057a", "validation": {"axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003510093782097101, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3d3493ba-7f57-54e3-94c5-34e37f52057a"}}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/e32a6f2a5dee7bf94613d1119eb93728c2fece3a45f1376a14e0fcae0619dd08.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8961bb7b9822d25fd9740551eccef03d4de9c37e5a3fa77d2f478fb41c4705d4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/fa13a91791f94fbb3576517c9ba62a7bc7a050ca246f54116da8cdd24a749719.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/aafe354428fe8f4b4a5786fa99a3314b16ac6da164c097d35fd959d91b4d6d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [181], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 52, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 52, "uuid": "af60e984-b386-57aa-9a1a-447684a59bd1", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.4901161193847656e-08], "seconds": 0.0018545021302998066, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "af60e984-b386-57aa-9a1a-447684a59bd1"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/6ec2fdbd40852627ad40f6e45bb1861784e13f7603e921fa695b407a6651e541.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [181], "selection_step": 52, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 52, "uuid": "998dd71e-de7d-52ce-b64b-fd65c93b2c8d", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.5513619473204017, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "998dd71e-de7d-52ce-b64b-fd65c93b2c8d"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/9b50d91c80b7193668ca425bc833a2ca67830e7940931f2eea811d7c06cfffd5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [181], "selection_step": 52, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 52, "uuid": "87441003-b95b-526d-b0f7-9e90148ebee5", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.34558157389983535, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "87441003-b95b-526d-b0f7-9e90148ebee5"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7f3d60ab4b9815d9a00abe6dd4d87bf56c50653ef4866737f18408e0a0a7d6e4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [181], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 52, "uuid": "04f8298b-720e-5f9d-be0e-4318bf74de22", "validation": {"axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0034822807647287846, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "04f8298b-720e-5f9d-be0e-4318bf74de22"}}
{"actual_requested_output_tokens": 107, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2fa8dfa782ccb71979b7e78b45b04993c59fea59271cf633354c733f27a64540.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/af59d6504a76e0e09c7994c45327914803ee7e1b15b417758e14ecce2491341c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6a332a1c9104697eb316448e98aaa24eb906ec902529dd4a53edcc9419e4260a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/e94d8ec9d55a8545c5891c9e09175b3da49409b494e9bacbccfbdaee082d67d5.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [234], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 105, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 105, "uuid": "99c44408-2bf6-5b21-ac9e-d9d49c5eab60", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018702130764722824, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "99c44408-2bf6-5b21-ac9e-d9d49c5eab60"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/7ee77b2d5eaa8d59523f4c4fd6a90e761b356ca210a1cd3eb96cf1c8a8f6df3b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [234], "selection_step": 105, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 105, "uuid": "900885d5-e628-566f-9f35-d67aa7a15924", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8859752179123461, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "900885d5-e628-566f-9f35-d67aa7a15924"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/d7793b41566da8ff92174b4fe317f48bd126dd0285883275d88cc1d866274b4f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [234], "selection_step": 105, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 105, "uuid": "f94c4eb5-7e50-5719-a440-040fa6dc1bc4", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3394766910932958, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f94c4eb5-7e50-5719-a440-040fa6dc1bc4"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 107, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 128, "label": "short_request", "locator": "requests[22]", "model": "claude-fable-5", "output_tokens": 107, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Shortest recorded nonempty-output request; sparse-history boundary.", "percentile": 0, "target": 128}, "session_id": "16c18d979235980e8021cfb0a0b3b386305a", "t": 1452.253, "uncached_tokens": 128, "uncached_tokens_256k": 128}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 52, 105], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/57d71f41692fe6a85ea2572ea2af77bf7a38d4e21377816da3182fc2cccdc327.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "5729f5afa5cdb1661ab1322da2559cdc1e4e03e927be5481169405ed4b4ce9b6", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_short_request", "scheduler_mode": "decode", "scheduler_sequence_lengths": [234], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 105, "uuid": "1ad1d4ca-c265-54b0-8bcf-123948e84b86", "validation": {"axes": {"B": 1, "N4": 1, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0035882792435586452, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1ad1d4ca-c265-54b0-8bcf-123948e84b86"}}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2944, "Nkv": 2944}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/a0c2808625627f1d5867bfcc25ee7419f018ba66425d689f6e7b7405d53525ce.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [2944], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "ec18845f-e1bb-5eba-9649-3d90d470ce9b", "validation": {"axes": {"M": 2944, "Nkv": 2944}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.1716800257563591, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ec18845f-e1bb-5eba-9649-3d90d470ce9b"}, "variant": "rows_2944_compression_0_cached_False"}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2944}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/5d8c3e4fdb71cec3d2ac6c09410eb5085007af7186be36f045e7e1a18e6d9ede.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/98ff14aef32cc15b3569c3fd0cd26f94ec87d4142f00423b731832b0445968a7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2b8a4cef93226682381aaf5a1f95dd30ff0a543068bbb6d025169e2ba3be244c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/c717e32a02b46c9a2dd4cf566674122cc6342fafcea2c9f7ad4e5d8103ed8b29.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [2944], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "01578ef1-d67e-5840-b8ce-c8a26ccd7515", "validation": {"axes": {"M": 2944}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 3.2186508178710938e-06], "seconds": 0.008137943223118782, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "01578ef1-d67e-5840-b8ce-c8a26ccd7515"}, "variant": "prefill_2944"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 17664}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/7b3cfbc03b72048cf44ed8a4039e786b73b76e95e4609b535301f8d9a5312ed5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [2944], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "630f85b0-9ace-54b6-88a0-08bb930237f1", "validation": {"axes": {"M": 17664}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 1.2196662919595838, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "630f85b0-9ace-54b6-88a0-08bb930237f1"}, "variant": "prefill_2944_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 17664}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/c68b608397eb24c9b125dad1eb246e8061b89dd9dfdc001f5f3e21f4a0a5afc1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [2944], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "acae606c-4112-548b-b870-cec7314a3f82", "validation": {"axes": {"M": 17664}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.5008114967495203, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "acae606c-4112-548b-b870-cec7314a3f82"}, "variant": "prefill_2944_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2944, "Nkv": 4416}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/0bacc9640c92375c4b994168ebb73235f44df4b1beabb763ef988c8f9a6892aa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [2944], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "6a0463ea-da74-5e52-a6f4-18548159142a", "validation": {"axes": {"M": 2944, "Nkv": 4416}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.16766805900260806, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6a0463ea-da74-5e52-a6f4-18548159142a"}, "variant": "rows_2944_compression_2_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2944, "Nkv": 5888}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/6417382a44a9c9038a2bac400f59f12b9a5d622134f9ef62ba2cd47e1d70d8df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [2944], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "52560299-af35-5351-a4f0-cdcf97141e87", "validation": {"axes": {"M": 2944, "Nkv": 5888}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.16742154397070408, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "52560299-af35-5351-a4f0-cdcf97141e87"}, "variant": "rows_2944_compression_1_cached_False"}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ecbb2abef863d2c678d2a73ea15c92c4da840f15f15a2f12bc555bb1a1b2f537.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/53274d3b8f9e5296f5391b03e24829abe888b0f5db1f2c03ffc91ba0586a8e42.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/3512952dd2bb124e7766f7bcfe8ff20f3a72a39ad2253b639e385a64a16f5ce6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/5a8a3b812929bb630e5d81e36c94f640f353d61dbaba4c9df0aa27d3ee9fdc02.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7a7cdbd4-6374-51f1-9f99-53298395c41d", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018858392722904682, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7a7cdbd4-6374-51f1-9f99-53298395c41d"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/2bfc386cf723d8edebed0e10705c122fb4b29573de51af6b14f934b41691b491.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "88e96974-46dd-55bf-a772-be150b017d9a", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8851861101575196, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "88e96974-46dd-55bf-a772-be150b017d9a"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/5cc11b258e00ce827774802b27b16de4798d35894d5266b4a48a76dbb5c8f687.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "a83055fa-d10a-58ed-adb6-de682564c8cc", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.33914247201755643, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a83055fa-d10a-58ed-adb6-de682564c8cc"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/3c4a0109de069cb70336437c4a360dbe30b205f46a5660ecbbb36f8721957fe9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2b87464912950073b063c001cfeda4bf4f6e105ba4d22fce2437368c298436e1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d715c1df4e5f698fbf31d7c76c7a47b5b34d44e359932d58f91f7014b22c249e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "2c055caa-e9f7-575b-9d5a-36376e19011a", "validation": {"axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004787247162312269, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2c055caa-e9f7-575b-9d5a-36376e19011a"}}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8a09524e6a0885e674491813771a97ec342439532070cef8bce961c86064b318.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/41667d74739c9325aed400cacf15357da10c9abd514c3994fcba4ea068d1beb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f67ef9cecd7a5614930dba5c720a9e6bbb60a1352992ae77d7395793f0e7e6b0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2db5bf2fae7e10ab42417743f036c4833075f1b405995419685d0a769645cb0d.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "2898ff25-4c56-50ba-ac84-59f2d0d11dce", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018871850334107876, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2898ff25-4c56-50ba-ac84-59f2d0d11dce"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/67be217d1f1b345b7acb770cb15b68678df3ddbce7a96616fa2611fc66760663.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "6c0c3272-75bb-5404-adda-77210a1a9ccd", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.882862709928304, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6c0c3272-75bb-5404-adda-77210a1a9ccd"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/9f4bb3d95910e0d13a3c06c3150702fc50658ab733b73b335145eabc7b092d81.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "0d2e28d9-5ff0-5df1-baa6-ce422d564244", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3401966840028763, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0d2e28d9-5ff0-5df1-baa6-ce422d564244"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7efc14437796e3a9def741b8ccfd23401c5d3667f5491736dbf67b3488f9a822.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2b87464912950073b063c001cfeda4bf4f6e105ba4d22fce2437368c298436e1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d715c1df4e5f698fbf31d7c76c7a47b5b34d44e359932d58f91f7014b22c249e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "b703462c-a066-552b-9a2a-a557dc484675", "validation": {"axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0037732706405222416, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b703462c-a066-552b-9a2a-a557dc484675"}}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/d82afffac00ef0f037bcfe89fd81a5aa8501383643f4ef9d4563a8e806359b5d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6b24db472c6f5ca2d7d0c6a38f4b6b014ac4e5a0dc18498f998ac9b76df0bf6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/5fcfafa715b296db1f98fc06e4ba1091e495d8d59bf1ad22fbb05076527c8900.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/1e22412c060d2e6de86dd06c132b7edbd50dd3fd3680347b61282959d8da45a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "dc6ed9ab-498d-55e0-a719-90f442212932", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 0.0], "seconds": 0.001883982215076685, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dc6ed9ab-498d-55e0-a719-90f442212932"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/5351c07b703e638b20c7d6996f101df7a16b127fb7994fc75773898a467184ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "7fcc5dc2-4220-5464-a732-5b282de66991", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8487191488966346, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7fcc5dc2-4220-5464-a732-5b282de66991"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/eed89bf1258cf329af1120082ad65c909893c07839f19a99b5e84f0ed3b98f0d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "c63b7111-d931-5b58-b45f-9a4e05244eac", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3383104531094432, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c63b7111-d931-5b58-b45f-9a4e05244eac"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/6c3a5f9e70234d34d9119bf98227965d1b5c1eb018cea6e4cd698a0bbc6313fb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2b87464912950073b063c001cfeda4bf4f6e105ba4d22fce2437368c298436e1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d715c1df4e5f698fbf31d7c76c7a47b5b34d44e359932d58f91f7014b22c249e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "63d076d7-258b-56e0-acb9-9d2094dfa87e", "validation": {"axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0036922399885952473, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "63d076d7-258b-56e0-acb9-9d2094dfa87e"}}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2611a26297b6b8a7b0adb4d2cef7c1a092013fca09cac8882c076480c55f1c01.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/aa4f10ef3d2ee096f139ab939dcf966423dc00a4e439aa5a8ca255b4f7471381.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f65205944f2852ab00b0076eef054e616596420f862a5fb59787defea278f832.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a9fd9d2d975cc65d1eda502c562c3130c39c7e715aef8f452bf80204c361e820.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "b94d7d4e-92b8-531c-a7cb-4e312a5dd352", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 0.0], "seconds": 0.0018848809413611889, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b94d7d4e-92b8-531c-a7cb-4e312a5dd352"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/c6c1d66c8d69b3e1a2cfde8ce79ae13348f892aec9663beebf5019ae700cbc5c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "0ba3e03a-7163-5ba9-b206-d7bf44d614c4", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5414832346141338, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0ba3e03a-7163-5ba9-b206-d7bf44d614c4"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/b7f05bb25649cb16ab8b635bddaba5c0f36e289b1738d1bb34344121648d062b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "1b416d6e-080a-5179-913b-8aa5707cc3f7", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3382181888446212, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1b416d6e-080a-5179-913b-8aa5707cc3f7"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/12407dc271b16d2ed0a6ba44d105332c0fd4d0c1bbe27e5538e9eb5d314a0bbd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2b87464912950073b063c001cfeda4bf4f6e105ba4d22fce2437368c298436e1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d715c1df4e5f698fbf31d7c76c7a47b5b34d44e359932d58f91f7014b22c249e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "5df6a182-3f20-5ae9-80fc-9dc2b9171d37", "validation": {"axes": {"B": 1, "N4": 12, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.0035967896692454815, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5df6a182-3f20-5ae9-80fc-9dc2b9171d37"}}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/634c5bb6d35b04bd9760f9bf9508fe40309bf57057a39fb8887eaf3ef99804a2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9b257c83fee44304bba5cf11cfcff63a9979a901fe449763f13fde87ca417d5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b1e442a7b0cf6eca71ed385537595f533e7827e4a327bb65abb6f972211042fe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/4a0fda3b914dd8479103eb1959bc87c69a6700780afa635495a3e39b34d920e9.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3850], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 905, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 905, "uuid": "3429bfab-7670-5910-9da7-b80768819de5", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018188119865953922, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3429bfab-7670-5910-9da7-b80768819de5"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/cb66ac4f0cd1c5175891393aaf6535b8959dbb01454a0d8a74785c0088566473.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3850], "selection_step": 905, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 905, "uuid": "68bca134-bc2b-5e14-9ab4-68e182920db7", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5399576942436397, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "68bca134-bc2b-5e14-9ab4-68e182920db7"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/084c707685328192664f58d19c64af058073664f9c61cf530648c508e9fde818.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3850], "selection_step": 905, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 905, "uuid": "626d1307-1493-5593-8079-d1a58b9008e1", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3377021881751716, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "626d1307-1493-5593-8079-d1a58b9008e1"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 16, "N8": 2, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/a516db75e2c91026984ae37283d8ca6f11d91b7567c7d5ac5762d3bb26dc01fb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d265caf29c1023c85155dffb8e74f26c777f56e3113ed62c77955e8411c0c16e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/73867bdd1fc6cf6a8e969490f87a8578cc0bf91702be620e48996d51b695bcf4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3850], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 905, "uuid": "67c1d03b-d10a-53eb-bf5e-34df0f9df57b", "validation": {"axes": {"B": 1, "N4": 16, "N8": 2, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003837081603705883, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "67c1d03b-d10a-53eb-bf5e-34df0f9df57b"}}
{"actual_requested_output_tokens": 1813, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/09ccd29f0829f65025190c659895407160ecff9b3acd63ea211290aba52e59eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/67c842f238502711c6a5adc42ccec5aefad2c7a0c51351b64ad2b17989cf8c63.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/21c20f4cd20758f90ac00e0362996652d00ef5c67ad1f3fb19c4a96a72f52c94.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2f6c9cb96b41c1b7f56409108ecf4d37389b65450d99fdc3474edf33cce33010.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4756], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1811, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1811, "uuid": "36754b68-5dc4-5c2a-9dbf-c57ca9a6c523", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018437537364661694, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "36754b68-5dc4-5c2a-9dbf-c57ca9a6c523"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/92401f14b7149adf55c04a902effbe92f95f3067fd98764403cdbb134c5e3c57.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4756], "selection_step": 1811, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1811, "uuid": "9d2e0dc3-9185-5307-b78c-7a8c325db34e", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5410367511212826, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9d2e0dc3-9185-5307-b78c-7a8c325db34e"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/36452df18abf87401ce997afb6874ccde11cbad2811c4c944a3846b3c67c0d63.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4756], "selection_step": 1811, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1811, "uuid": "5bac813b-88eb-56c7-b06e-312a7ef55e26", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.33674867870286107, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5bac813b-88eb-56c7-b06e-312a7ef55e26"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1813, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "label": "cold_main", "locator": "requests[137]", "model": "claude-opus-4-8", "output_tokens": 1813, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "main", "selection": {"field": "input_tokens", "note": "Main-stream request with no completed-request prefix available.", "percentile": 50, "target": 2944}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 63769.496, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 19, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": null, "capture_decode_steps": [0, 1, 2, 3, 905, 1811], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/4d6f98e34f23db26f66aba205dbf952f4f17e92ca65bcbeb146d8a9f0c7503be.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "29f4cb73ab5e5b4efbca25f5c100876d73d535f9b7ef3856b82762f3758891e8", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/27b024bab8e0282f4d4a168257a72277b908f11d284e5eb312bc127b9e1f2c4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f07204d57b7e50c1a71e4c5f5b25bc5858c3dd21ef9fcfb12f9a820763e89677.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_cold_main", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4756], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1811, "uuid": "ee56d2be-7be4-5091-a4e6-d827a6c6c4e3", "validation": {"axes": {"B": 1, "N4": 19, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.003902407828718424, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ee56d2be-7be4-5091-a4e6-d827a6c6c4e3"}}
{"actual_requested_output_tokens": 90, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 448, "Nkv": 575}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/2e9515ccea2d768065a2df3fbf622d477617ce2c04d40ee5c323196fd25a0930.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [142016], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "d1602ca9-fae0-56ef-9bb2-a408fe19a22b", "validation": {"axes": {"M": 448, "Nkv": 575}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.029641964938491583, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d1602ca9-fae0-56ef-9bb2-a408fe19a22b"}, "variant": "rows_448_compression_0_cached_True"}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 448}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/31fc42268045ee1e94618e370c0ac5602e9cb8b05f42b0cfe6ddd6055b092b0c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/679a3b46f81cbf1063e07f9259961c967556d9c4b8f7b410e59c65bf5f7ac659.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a9fa3b23fcbae3c06c3971f0f55d8237b43c0c98b378132eb5b90242f78d34a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ab3c19d5061aec6ec16d200ca91d4beda78f4937e22d6953857dcc5f59b3adba.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [142016], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "3e29c0bc-7220-502f-b98f-d12f5720076c", "validation": {"axes": {"M": 448}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 9.5367431640625e-07], "seconds": 0.0030147982761263847, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3e29c0bc-7220-502f-b98f-d12f5720076c"}, "variant": "prefill_448"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2688}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/3fbf5e26e4521d9991708f9fb19f4195411b7cc38993ff921c4a485d05f81b5c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [142016], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "669309b6-8328-5a7b-a2ff-dcae81dfd4f8", "validation": {"axes": {"M": 2688}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.71800921484828, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "669309b6-8328-5a7b-a2ff-dcae81dfd4f8"}, "variant": "prefill_448_cached_True"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2688}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/721e2d71ad17d9562ecb9b29fde53181a1269552fffe6c1a661f0e0ce1409883.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [142016], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "9972c26f-b34d-56a2-a043-a50684289467", "validation": {"axes": {"M": 2688}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.43521131109446287, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9972c26f-b34d-56a2-a043-a50684289467"}, "variant": "prefill_448_cached_True"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 448, "Nkv": 71583}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/5aa0f0f71df781da73ddfac2f89060dafd5e25c31704012443ab3ab4b32c628d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [142016], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7d138df7-8e65-541a-9ad1-d572d5131b5b", "validation": {"axes": {"M": 448, "Nkv": 71583}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04861454013735056, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7d138df7-8e65-541a-9ad1-d572d5131b5b"}, "variant": "rows_448_compression_2_cached_True"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 448, "Nkv": 142591}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/16be1cf36a53699d6833d31f11c3c0fed1949831db510d9213d77c0138126018.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [142016], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "495000a6-c8b0-5057-99f4-88ed4a96a4a1", "validation": {"axes": {"M": 448, "Nkv": 142591}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.08206645119935274, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "495000a6-c8b0-5057-99f4-88ed4a96a4a1"}, "variant": "rows_448_compression_1_cached_True"}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/09ccd29f0829f65025190c659895407160ecff9b3acd63ea211290aba52e59eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/531586492d553c2e5470fea35d1a8e71c1b51640c9eb7cf4d382bdee15b35fe7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/21c20f4cd20758f90ac00e0362996652d00ef5c67ad1f3fb19c4a96a72f52c94.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/124078be107aa872230628cf9f1a4bf712639dde8284a56d8e83a208cf77cd93.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142017], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "1465f204-f3a6-5d8b-9d52-b0fa13447170", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018743081018328667, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1465f204-f3a6-5d8b-9d52-b0fa13447170"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/fad04446054828a3cef9ece6aff7d0241474b0a4eb63689b4cff5038f331f5a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142017], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "42076ff8-4ca8-51a7-acd6-22701e68ff25", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5414900947362185, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "42076ff8-4ca8-51a7-acd6-22701e68ff25"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/479a2dba0999aae37cce8d64d834489a278533abaaf9fa7fe7aa9433a2c080cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142017], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "94166d72-e40d-5565-b9b2-db03047bd74f", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.33660757867619395, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "94166d72-e40d-5565-b9b2-db03047bd74f"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 277, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/762550f7a0d62818904f9e9d03bf7501461ed421c5e8dc946d3251818721882a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/477c74a8854f9cb323e4c40c2c3e25b13832aec507be35c33d812da2288737e7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/627ae435e59b7294f04d8545f05db67cfa01059e3e470743edf6b1156774495b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142017], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "ce862982-589a-547f-8b47-e68851a8566b", "validation": {"axes": {"B": 1, "N4": 277, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.006942179054021835, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ce862982-589a-547f-8b47-e68851a8566b"}}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/3f35a46e1b71132aeaa26b11d9f8669870077ff92d60304db890f1071634252a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/703fc47cd09c1854850618732b90bcca5a8d5d830a4f6c9c6e1e03aa28c27a10.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a1402d76991ae1204549d6a7eb6d6d246b2f775991166f6daa50849d38f7c38a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/15608b0c826f90f4ac9d43ca49f675270843db6b910ce9daf5a23c7a87bda982.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142018], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "32358a16-c098-54c5-b08f-2163b985863a", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 4.470348358154297e-08], "seconds": 0.0020304671488702297, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "32358a16-c098-54c5-b08f-2163b985863a"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/49cc025ea813cde0c10ad4a4975379ef399c942158fdd176c0e7939f262e108c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142018], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "83936513-cfbd-5365-b669-4af3541aa223", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5406065662391484, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "83936513-cfbd-5365-b669-4af3541aa223"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/5407919f93a422116875fa0a8c24ab2b93e1aeed07ed94fb961f932e2aaa1676.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142018], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "73b67c4d-01a7-5566-ae5e-07adfad25dc2", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3359463354572654, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "73b67c4d-01a7-5566-ae5e-07adfad25dc2"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 331, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7ae5a3c5b106a4e7a9b372f1d2317d8da960ca152c84805528ea2e68ca81bb4f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d741354eb5ed44132c55bb53a3fd21077042d16311b653c01ec2325ce9ea52f3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/627ae435e59b7294f04d8545f05db67cfa01059e3e470743edf6b1156774495b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142018], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "f907c5dd-eac8-5661-a36a-56aa7a7aa6ac", "validation": {"axes": {"B": 1, "N4": 331, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.007241962011903524, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f907c5dd-eac8-5661-a36a-56aa7a7aa6ac"}}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b7100fd06873521c15063e004041dbbb5ab115d41e088e104cee155978691689.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/bc0019a176d85791d2541854684221862122180d104b3062b1ba0d2eb7542176.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a6132df243d5e1e5089ec529cfaf84b6118a142894660255316af37cf9567a8a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6a61e6220c02044b23e773d7331ee02655c05e23abb9a7055e020a1eb20eca5c.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142019], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "a89c8fd6-0750-5bee-bdd9-4a04d254cba9", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.4901161193847656e-08], "seconds": 0.0018712240271270275, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a89c8fd6-0750-5bee-bdd9-4a04d254cba9"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/2f4dc6e6f2fb9e2485019da36bd339c44e6f16cad6a483f2e235ae4672540811.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142019], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "448d2841-0915-5817-b45f-0b0accb0f09f", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5402844678610563, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "448d2841-0915-5817-b45f-0b0accb0f09f"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/39643d5869435e494582a3713560e2c3cc632af93b21fc616085f463d2177ee9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142019], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "c5e9467e-5252-5532-aaa6-28a4e74a6789", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3366624996997416, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c5e9467e-5252-5532-aaa6-28a4e74a6789"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 345, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/db6d05c680d878ae66bd14d7eca06b58227b6087a5be9e2b876eb5bff57d0ffd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/8c5b2304cf8c552e21d25210358fd1916d95ba6510c5d3f338f67bd178d240a0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/627ae435e59b7294f04d8545f05db67cfa01059e3e470743edf6b1156774495b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142019], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "2aecb81c-d898-53d5-8556-3d0991d65651", "validation": {"axes": {"B": 1, "N4": 345, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.007075177039951086, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2aecb81c-d898-53d5-8556-3d0991d65651"}}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/092125054f5770a954bd7ad7d3fd694693f4b7c5c25fb339ffecd4223772a89f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/73066f038df87216b0dd08d84218757d2aa84d4beaeb51ca4eeb075ed43f667a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/58d632efb31ec216cd82bb4c9ebeb11619463ef7851d9a1874049f8f36000352.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/655e0d275def327f90477d4a263637222b207fbbb5d5a066ff44f7bd737e9ec1.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142020], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "26024620-0ef7-5520-8f45-41cadd789175", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 7.450580596923828e-08], "seconds": 0.0018882136791944504, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "26024620-0ef7-5520-8f45-41cadd789175"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/00dc8e682857a357b5cb113b5bd2ecf711e4e009deaabefac7ad46f1f1cc32b5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142020], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "1d7e47ca-8d42-5103-a323-4baabee79fce", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5398206459358335, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1d7e47ca-8d42-5103-a323-4baabee79fce"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/095daf71d11aced296aa26e3263a5d636dd496efa299718d42355151c244ff9e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142020], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "3ba9cb47-0488-53a7-ac2c-21920bac8a55", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.40999557124450803, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3ba9cb47-0488-53a7-ac2c-21920bac8a55"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 350, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/aa34027ce19d1b5fd9b111ab9815d52e924063790b3bdd77de783f7fe6483589.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/982661c49712bfb972d47d3745f7bfaa9a70f3fd78ccec21fd483317224a57df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/627ae435e59b7294f04d8545f05db67cfa01059e3e470743edf6b1156774495b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142020], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "d2090849-b97e-5099-ab49-e5b472e4cc56", "validation": {"axes": {"B": 1, "N4": 350, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0067698098719120026, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d2090849-b97e-5099-ab49-e5b472e4cc56"}}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6d429a556db57aa1339ef11a90e4351d93168c625b0a70826e506c10d5f22319.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/10fc931b26250f3875a489a56de3b4231954a70ca5d0e100926b8eee25f495b3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/23ea213014268f865ee574bdb75aecc5946545d58c3fadfc3e53c259724b2a8a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ec764eca7772d53199ec0489dadfdf5296b90f02f4def3ac55c97a0e9d04455f.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142061], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 44, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 44, "uuid": "693a1e16-0aa3-50ad-a4a7-f7ad88f2ecd4", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.4901161193847656e-08], "seconds": 0.0018469691276550293, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "693a1e16-0aa3-50ad-a4a7-f7ad88f2ecd4"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/65206f1fb0f4e5b68da4b8550ad3945901c87c4c9a45f47a34a153182a7d9784.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142061], "selection_step": 44, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 44, "uuid": "9e4664d0-3249-543f-88b7-da51dd8a7bff", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5383333088830113, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9e4664d0-3249-543f-88b7-da51dd8a7bff"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/94d6a9ec2fe888116e1b24d9fca75f6d2acb842bbc6cc6e149a9ec1d6bc08649.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142061], "selection_step": 44, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 44, "uuid": "8a77e7e0-79d7-5650-828d-2739836e20f1", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.4002616498619318, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8a77e7e0-79d7-5650-828d-2739836e20f1"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 402, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/13259280168ddf8e913bf47e47dba00417d263046a052a1df51a20dbf9cdc1b1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/b89a7d32aa5dc84f736377d8d13c816a56f35a7515fd9df8df73bcb4450730ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/627ae435e59b7294f04d8545f05db67cfa01059e3e470743edf6b1156774495b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142061], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 44, "uuid": "484ddd41-b31e-55d1-9f15-9ff33653f4ed", "validation": {"axes": {"B": 1, "N4": 402, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0072348639369010925, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "484ddd41-b31e-55d1-9f15-9ff33653f4ed"}}
{"actual_requested_output_tokens": 90, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/11da69eaf8fefaab4fa858f5f3d2e3a31a27c3d2570c7a589bfc7b01786a085e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b9fd9bcef3855ece2b21f284a09335687b7c3015189164d11c518ccf3eb2a117.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ef4bf00a5b167d6644e3247d4c6a8ba0dbd2c1649ea4005a19da42480aba063e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2d2875093a95a79836259fbaf778c7d70884a4371ec8836305d746615f4ddca7.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142105], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 88, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 88, "uuid": "68e00c17-4a02-5fe0-ac99-f0423af525f7", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.1920928955078125e-07], "seconds": 0.0018689977005124092, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "68e00c17-4a02-5fe0-ac99-f0423af525f7"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/7dcd475cbfc559eec256077d51eb7a564e55b69562ea804805d36514186b5f5d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142105], "selection_step": 88, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 88, "uuid": "6ebc17e6-ac78-51b3-99c8-183fe2190433", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0009765625], "seconds": 0.576311593875289, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ebc17e6-ac78-51b3-99c8-183fe2190433"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/ad8f02c0f6a4c101c7917b1fd4638e47bc2561d97a350d3bfb19a036ccec35b8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142105], "selection_step": 88, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 88, "uuid": "7d1d7a29-8b86-5580-b237-9e5194ca9a6a", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5318765640258789, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7d1d7a29-8b86-5580-b237-9e5194ca9a6a"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 90, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 142016, "label": "full_context_p50", "locator": "requests[452]", "model": "claude-opus-4-6", "output_tokens": 90, "reusable_prefix_tokens": 141632, "reusable_prefix_tokens_256k": 141632, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 50, "target": 142016}, "session_id": "0470d446a4514dfe0c6ad0be92853bd13287", "t": 113074.811, "uncached_tokens": 384, "uncached_tokens_256k": 384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 180, "N8": 2, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[451]", "capture_decode_steps": [0, 1, 2, 3, 44, 88], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/6332feb5702e91e78917c8f10c44af16d94a0a07cbda7a546fde95e1e3097023.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "e262c52fc5a556a94e421a9cbbe377b90692e259758d39a27496d2929356cf84", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/b6dcd3df671f503fa224df6c2b65f7b5b9b10bc94c69f9f718eb95b5dbf9b0ef.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2cb6d4b7131c3d3129a07423fbcb290f03b47c72ba96b2dc271b844b95bcfa85.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [142105], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 88, "uuid": "680bd5dd-94fe-5dcb-84d0-9488c1db10d7", "validation": {"axes": {"B": 1, "N4": 180, "N8": 2, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.005677140783518553, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "680bd5dd-94fe-5dcb-84d0-9488c1db10d7"}}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 4223}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/057baf4ac56eada3f421e3b654715f92f567b4ac36ec23076ed55c2c288b9dcd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [546048], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "d1c89357-aa68-53ac-b13f-62d900007022", "validation": {"axes": {"M": 4096, "Nkv": 4223}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.23019856214523315, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d1c89357-aa68-53ac-b13f-62d900007022"}, "variant": "rows_4096_compression_0_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24576}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/b7ea650661faa21128e99ae8a360d1a6c8f3b6c2db71898712ba2bc6bada3fa5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [546048], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "6e152904-15af-506f-a1d7-c4dc0b73f749", "validation": {"axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 1.1952670449391007, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6e152904-15af-506f-a1d7-c4dc0b73f749"}, "variant": "prefill_4096_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24576}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/97d3501a043b033e123357b0360f2d08950c0e33dd6ee6e4cb19878a30be73fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [546048], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "c6405726-c01d-5e0f-b77c-fc20a20c93ab", "validation": {"axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.7681796588003635, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6405726-c01d-5e0f-b77c-fc20a20c93ab"}, "variant": "prefill_4096_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 277247}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/8998d81445f99955cea4890a5fd74e14619fd0770a5e0e19c4bb4e66a308bff0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [546048], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "1452f1b3-6d7e-5490-ba5d-ca2276922417", "validation": {"axes": {"M": 4096, "Nkv": 277247}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.30069290520623326, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1452f1b3-6d7e-5490-ba5d-ca2276922417"}, "variant": "rows_4096_compression_2_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 550271}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/ea8c1db1cde54e4b71dd9162c3907e0d4f224226f7a1a598d87caeacb4a10e88.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [546048], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "c0332d0e-dddb-5656-a8c7-0ced515dbbfa", "validation": {"axes": {"M": 4096, "Nkv": 550271}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.8619707771576941, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c0332d0e-dddb-5656-a8c7-0ced515dbbfa"}, "variant": "rows_4096_compression_1_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3456, "Nkv": 3583}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/9d6322037b8c795d5ee50cf7e3f27b2dd2d0e3093774361004c6bb0df96720ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [549504], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "c315bd09-f7c1-5174-bed2-66b2e2b8d7be", "validation": {"axes": {"M": 3456, "Nkv": 3583}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.19409079710021615, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c315bd09-f7c1-5174-bed2-66b2e2b8d7be"}, "variant": "rows_3456_compression_0_cached_True"}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3456}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7543e9bbce414d64a0732dba69ce43f0ea7ab43f1139f01d2dd1916385e922f6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/57f3f5e2b28bac04085691a0a21c713f56790b54b280893e59f2ec76f752ec6c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6ebf6770d66c58ac6fa2363663f29caa8d9f1e8be8b821863c9d2e50c415d564.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/1e889f9db33dc5b24ea22256ff74096ba0b82de45518c0ceb89eb7f3bfe1d08f.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [549504], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "de8359bd-2cfd-5d2e-8c9e-d78390c9b27b", "validation": {"axes": {"M": 3456}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 3.337860107421875e-06], "seconds": 0.008849332109093666, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "de8359bd-2cfd-5d2e-8c9e-d78390c9b27b"}, "variant": "prefill_3456"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 20736}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/996fe8e28b8a86d9c1f1867cc462932eab9d76bc83cc922a2685d83fb0cc454e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [549504], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "a53c635e-3cfe-5694-97e3-b0572332c136", "validation": {"axes": {"M": 20736}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 1.1965465568937361, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a53c635e-3cfe-5694-97e3-b0572332c136"}, "variant": "prefill_3456_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 20736}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/3c82585ec4ad6e3213842b4f8763d5a2541b688e8fc947613a5d8f18da3e92c9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [549504], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "93be97fb-e453-5751-994c-2f88cbf6ff30", "validation": {"axes": {"M": 20736}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.7529570069164038, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "93be97fb-e453-5751-994c-2f88cbf6ff30"}, "variant": "prefill_3456_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3456, "Nkv": 278335}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/d3cc24110494e5527024a63fb71ef9e595d30064dea53c65265cd31f0c7e3e8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [549504], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "d22bc34f-f49e-5883-a140-f1c5f6bbb8ed", "validation": {"axes": {"M": 3456, "Nkv": 278335}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.24783538607880473, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d22bc34f-f49e-5883-a140-f1c5f6bbb8ed"}, "variant": "rows_3456_compression_2_cached_True"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3456, "Nkv": 553087}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/08b6a44f258777ffdf4677bfc0016845b942a995ddb9e4b983905ba457076025.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [549504], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "5e5ad68f-a0a3-56e6-a0c3-065765feb8b2", "validation": {"axes": {"M": 3456, "Nkv": 553087}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.3047167588956654, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5e5ad68f-a0a3-56e6-a0c3-065765feb8b2"}, "variant": "rows_3456_compression_1_cached_True"}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b88a18b26979aa1a144dbb5bf1c92e630b6afd9c04cea9a76059d9daebc00bb8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/671453c01e9119fcc29636eb2d54b5ec6d11dc3c168db63006434db5b07289ce.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6fe606518897335d43fe744ea0eae34ef1ba71dbd26e3cdc5d182c35d9044030.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b945544d94def7d390c0be45633ffe15c4d5820515437ba7cab8ec98a5aa4c15.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549505], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "2032b687-a0fb-5150-a045-28fdfaea612a", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018634181469678879, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2032b687-a0fb-5150-a045-28fdfaea612a"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/4e9f5881ef740f4869a7d45e56ba53e3d6a926afee52ccd853b60e95fe4ec74e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549505], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7155ad72-d765-5626-9fbd-b1c55d02ef6a", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.876852915622294, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7155ad72-d765-5626-9fbd-b1c55d02ef6a"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/13dc8b8335236d1a531f0552dc5745a8b17a3647f65a5e66a1110fcbc9f474f7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549505], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "819e56ce-0e9f-5d80-ba29-2b4b66f02f1d", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5329804020002484, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "819e56ce-0e9f-5d80-ba29-2b4b66f02f1d"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 369, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/0397b419e10c3671e7205c93fe56ce8eb081370ed3e253fa11240ec73c111b12.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/8c1fb852b1d984db8a857b374e277719ee129fd98ef9db5b4dc88295dbdc91f3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/43fa684a1d8e0bfc46e8fbc3e82a85a2c0ded168db4c0554bc8f95b007014cf9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549505], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "b3caf995-3e21-582e-aeb0-c4ccf7c54692", "validation": {"axes": {"B": 1, "N4": 369, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.006881288718432188, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b3caf995-3e21-582e-aeb0-c4ccf7c54692"}}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/452bed498a4492979528d657735f11b0aa8606a9d9633acc1c537ef9327508c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/49c3abb766f90072bc607266edde3fcb24b663a52a458d25bfe0b2dfc8476ba0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/872c33df706bd158a852d89c74321f841e5f3b64a4ccdc80212418efebb77172.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/17b4ca29db80556985ae2d2e435d3e64a39a4341d4a26d647d3e34ba6ed7f572.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549506], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "de10ba0c-dc65-5e21-9b85-b17198b0909b", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018859980627894402, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "de10ba0c-dc65-5e21-9b85-b17198b0909b"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/476ec669cddee30aa5b44b83b6b41f1c120c9debe41588de1d61cf9f7d5cb1cb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549506], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "6a4af5a8-f47d-5751-9fe1-a984e78bbca5", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8794198879040778, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6a4af5a8-f47d-5751-9fe1-a984e78bbca5"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/18a1ece55f8d06f4ffcd58b7314e09804e82dd07e73614a5cb0b80aa38f3901f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549506], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "1ef8324f-96bd-5a0f-bb09-6754b31ab356", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5345803159289062, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1ef8324f-96bd-5a0f-bb09-6754b31ab356"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 388, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d1ce3d8c33a274908aa69eb9d482166f6ee1a3593d1cbb93f708674ce3b1fcb3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f786555415eb5fb2e376f57693883af7602b24814da9bd3f81424e9424e9eb27.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/43fa684a1d8e0bfc46e8fbc3e82a85a2c0ded168db4c0554bc8f95b007014cf9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549506], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "6184cc39-62e9-5827-9a9d-b8bf1f01226d", "validation": {"axes": {"B": 1, "N4": 388, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.007385804317891598, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6184cc39-62e9-5827-9a9d-b8bf1f01226d"}}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9fe6c90a35151e65ebdc069a86b6064d7775fd9f494c150ec3dbacb9560f5a0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/75d7e1bc4e1302295cd0841822a9803a1d2765919f5f57eb66a80edaa2100c57.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/03505d5680fee2ee2e158f102ed1bd9f42226dafd4e8c11e4e7611c537a2d0a4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b241e4e05bf52529a13a111a188bd3b826fc6af20cc5a5bce3a80c22d6d71d68.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549507], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "ee51716a-39f7-5dae-9967-73e1d73ed8a5", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018774350173771381, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ee51716a-39f7-5dae-9967-73e1d73ed8a5"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/e89ec0981db984bc9f79f36faae81c7482efd69cabd51b856ae682d171b6bcf2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549507], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "025dce40-e338-52be-a06d-c702cb4f4eff", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.6839453438296914, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "025dce40-e338-52be-a06d-c702cb4f4eff"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/eeee74431f12d103f1ae94bcb6a0c60863d6f0d67c75a7e98c1e8dc63283aa1f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549507], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "8813329e-c415-5901-baa7-b9f423f06144", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5335805821232498, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8813329e-c415-5901-baa7-b9f423f06144"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 448, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/5ea07399fb2809704213b49e477a34006908b308ab72e151baf8b43daa2adb05.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d6c58b288c10f6e25577024887246c2657e7dc5b8a3790059d216d4b77199308.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/43fa684a1d8e0bfc46e8fbc3e82a85a2c0ded168db4c0554bc8f95b007014cf9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549507], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "1cdf60bc-edf1-52ef-a731-dc5484a307f5", "validation": {"axes": {"B": 1, "N4": 448, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.007474483922123909, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1cdf60bc-edf1-52ef-a731-dc5484a307f5"}}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/587d50a1181bfc2e115d0d7679fe35343677b438ae18b8445097848be92fa88d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/86625bc365f40869284858790d370881c24d2ed78814af42969288cc92276cee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f9d27fe6521755633461dfe99c226446ec10aed05cd3f7df59cf53a6ea82347b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/3aa987c47d80e5080684dd4ab09baed9c3b807fc761551acdd3c6f29e0703a2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549508], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "7a8a85fc-f158-5260-9b7c-f03d75210337", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0019132322631776333, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7a8a85fc-f158-5260-9b7c-f03d75210337"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/e0f2db5c6d63f2099c0c0f335dd5861cfaf99bf7d27e8e4da3b4ac2addc7a1e4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549508], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "d0761c7b-6039-526f-a122-9b9e39f94b0a", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5393422758206725, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d0761c7b-6039-526f-a122-9b9e39f94b0a"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/09bd9c9a9a58903fecbfdd12c9fb123e1f120750489f2716fb6d8ecd4a800438.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549508], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "b6c96f86-5c08-5382-ad66-482a71b55b13", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5224500061012805, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b6c96f86-5c08-5382-ad66-482a71b55b13"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 441, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/af10addf0e19c685a1c78ca8a96e5b7ae33d5300b620c7fd1ce8f192bfc69fc0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/8d6375a2306a4ab79ecaa5443ca71b2a36297afcc41c7f9a544cc66164c1fa6a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/43fa684a1d8e0bfc46e8fbc3e82a85a2c0ded168db4c0554bc8f95b007014cf9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [549508], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "c89ae7dd-6f4f-5366-8098-a7fd429fb63d", "validation": {"axes": {"B": 1, "N4": 441, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0076981522142887115, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c89ae7dd-6f4f-5366-8098-a7fd429fb63d"}}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/44d52cc60514bf928fd8d8d8ad8da816103b2bc754ba51261802d00d7087ac2d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b69c8ddb2d1363ad3185a3fc093203bbb8eb4346d5bf1ed0a1c42ff3fc65648f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/d9cf2a59a1f40e05bc61cfb389c41a977ee7294e2021d7e684ea4211bd2bf281.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b44fe0ccc9eb232715ed1e7658cd8ffdeb062f1b39535abe1d86dcb2c264a59e.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [552886], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3381, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3381, "uuid": "d52e2810-bf81-5ec5-a053-98a5d182a485", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018911552615463734, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d52e2810-bf81-5ec5-a053-98a5d182a485"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/46c10ee165b7b14dba209e5b6c246a00ca7393eecdf301049e4b1d813d750e15.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [552886], "selection_step": 3381, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3381, "uuid": "38b9f50d-a08b-52a3-a74c-bebc70056a55", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5385668869130313, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "38b9f50d-a08b-52a3-a74c-bebc70056a55"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/58ca62f583add82b057c3939488006f72162b2e1e75c899f2d6f6440bc66fbb4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [552886], "selection_step": 3381, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3381, "uuid": "f0f147c8-68cd-585e-a737-a1da8a53662c", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5277281040325761, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f0f147c8-68cd-585e-a737-a1da8a53662c"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 372, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/b7446a392a91c18e6e2807e6822d4a7a450df70d531d273e62aafdc977c3fa07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c259b7e48e226390555b8c79c52b5ab64b32fa536b090f7d31a7bfd2c292e068.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/dd4f3fe39c0fa38a5de8c79eb50d26084b6fa371b0496caa5831042dd3084a21.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [552886], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3381, "uuid": "f1361c58-1268-51bc-9952-5b6fc88b9419", "validation": {"axes": {"B": 1, "N4": 372, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.006781119853258133, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f1361c58-1268-51bc-9952-5b6fc88b9419"}}
{"actual_requested_output_tokens": 6765, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/400fe4c46a02e4b156dea951bc97354ebcf6ebaa35f6fb0a38290166e0aecb06.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2d0a0c1f6559d7525675f4f2babaf4f71019dca88e376ef8796aa4b928752e8d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/093844abca9c400c9b1d582732b3b0e6dd6f1d178861e3275a56a54bac5628d5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7b5d7cc57c7ba685d037f8ddd0324f90c10ea925cf8a2a65776c272d6ea6a180.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [556268], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 6763, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 6763, "uuid": "d2df577f-b8b5-5fb3-931c-54713f9be07d", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0019130720756947994, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d2df577f-b8b5-5fb3-931c-54713f9be07d"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/53a57dd38899f4d36fca9da13688a2cfce6186b5bbeb1af2718ce5576dc81827.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [556268], "selection_step": 6763, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 6763, "uuid": "95953dcb-b094-5569-9867-53b56eb7fb4b", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5377042600885034, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "95953dcb-b094-5569-9867-53b56eb7fb4b"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/dbce240f22d5dac2b0d5db6e544f273c7b80f2f3a3cfce44c28ec12917b8d3b4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [556268], "selection_step": 6763, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 6763, "uuid": "671e5f78-5429-5d4f-94af-24627c22937a", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3114455109462142, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "671e5f78-5429-5d4f-94af-24627c22937a"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 6765, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 549504, "label": "full_context_p90", "locator": "requests[202]", "model": "claude-opus-4-8", "output_tokens": 6765, "reusable_prefix_tokens": 542144, "reusable_prefix_tokens_256k": 252608, "role": "main", "selection": {"field": "input_tokens", "note": "Observed order statistic at round((N-1)*p/100); exact recorded request.", "percentile": 90, "target": 549504}, "session_id": "21cde366f5bd2f0ff2031b0538496c7e2a05", "t": 10289.451, "uncached_tokens": 7360, "uncached_tokens_256k": 296896}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 457, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[201]", "capture_decode_steps": [0, 1, 2, 3, 3381, 6763], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/68e405c45f2767e885f3d77898cbbce68a4a2ccb6e99bd2897ed5d192afb177f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "58fb3d2530a18e97702acd19d1eb3cdf8c663f49a4024e7902d38a3fcef700d5", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/821a761023b4ad34649c0fb2421f5b5d8517b1ac1d99e5bb89c15855bbda2d65.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/43fa684a1d8e0bfc46e8fbc3e82a85a2c0ded168db4c0554bc8f95b007014cf9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_p90", "scheduler_mode": "decode", "scheduler_sequence_lengths": [556268], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 6763, "uuid": "aee264c6-9720-5a03-ba9e-f37fa8b448b1", "validation": {"axes": {"B": 1, "N4": 457, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.009693556930869818, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aee264c6-9720-5a03-ba9e-f37fa8b448b1"}}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2432, "Nkv": 2559}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/f4d5cf25fe219e3b4fff39f3cdda881e4a0d49e62f9fae3c28df320be2407871.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [989824], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "8a834433-bbcc-5122-b239-be8569f77d02", "validation": {"axes": {"M": 2432, "Nkv": 2559}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.18988473201170564, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8a834433-bbcc-5122-b239-be8569f77d02"}, "variant": "rows_2432_compression_0_cached_True"}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2432}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/e06944592304ebe2e9f1e72f5cce6c6dfdd21e7216fe6c3e2dee72ae3a0647da.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ae004ce47260c26518607c66ada50e743a52b0bd9158d58b3b214509bc3e8953.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/865bda57626b56ec75f993c9e1ba7702478bd83fe641d93c08283c15eb443504.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/d3beaf1af137bc48ddf030ed9cd9f39b8575880b1a16751894f667ae861fb442.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [989824], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "6fd94728-92b0-551b-b4da-44400748dd68", "validation": {"axes": {"M": 2432}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 3.2186508178710938e-06], "seconds": 0.34629026986658573, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6fd94728-92b0-551b-b4da-44400748dd68"}, "variant": "prefill_2432"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 14592}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/9209deabd296cef7f020391c422bbc7a19b17d47edbe91a97d18832b2e41ca4f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [989824], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "2718713b-be58-55b3-aeb5-e3a38cce6f96", "validation": {"axes": {"M": 14592}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.7905320059508085, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2718713b-be58-55b3-aeb5-e3a38cce6f96"}, "variant": "prefill_2432_cached_True"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 14592}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/e1d1b3fd9b31a86fa8489c8a90911af3124c1173ce8b375b4a3ee1ee4948e719.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [989824], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "852cde60-0529-5e42-83ce-0ebfe07a7c0e", "validation": {"axes": {"M": 14592}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.47346740402281284, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "852cde60-0529-5e42-83ce-0ebfe07a7c0e"}, "variant": "prefill_2432_cached_True"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2432, "Nkv": 497471}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/d9dcb611421dc20fa83e7f9a3f6d6d14a85b90205c84e5150ce1aadbf559cfb0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [989824], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "762c5a7d-7dd9-5e91-a328-06c62d9f471a", "validation": {"axes": {"M": 2432, "Nkv": 497471}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.2923034420236945, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "762c5a7d-7dd9-5e91-a328-06c62d9f471a"}, "variant": "rows_2432_compression_2_cached_True"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2432, "Nkv": 992383}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/2e52d1b85d76c5d98ff3f8de5ba9ed7928e5e0f952c153468778595e2fb3683f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [989824], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "42f3b7ee-5af7-5dd1-bf58-71493057d3f5", "validation": {"axes": {"M": 2432, "Nkv": 992383}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.3822329668328166, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "42f3b7ee-5af7-5dd1-bf58-71493057d3f5"}, "variant": "rows_2432_compression_1_cached_True"}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/044c34a6c301f1014ec89450c7ea729bf312b11f7a2014b354cda843e527d8be.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2a03b86b6414c1205aea55f30c943e8988a75c9ace40e445ac0ead3fe011dd27.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6abc1a56e92245fd090e21c2a17e7e5dda6516a98108d9829270e8f6a1825ff1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/be80332ab064458e5a240a01c26b790413ad6eb90dad0b443c29b1157112ba49.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989825], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "288e8f54-fe3b-5db8-8133-df67f745a868", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.4901161193847656e-08], "seconds": 0.0018673227168619633, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "288e8f54-fe3b-5db8-8133-df67f745a868"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/cf449316695599c75b7b2ee69cbb133c9dbb9f022175478bf30b6a6b6fbede4f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989825], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "fdc0d189-a357-57ee-b045-c08626f15054", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5379030527547002, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fdc0d189-a357-57ee-b045-c08626f15054"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/323e1b7cc8b68ed487d74a5133f9220765ed192716b6ab11b7327abf177ae9ce.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989825], "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "3ff114c5-d974-5bf3-8502-c9c346627f67", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.30503664491698146, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3ff114c5-d974-5bf3-8502-c9c346627f67"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 391, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d5515e0ae7b6e66ca2faa1758aed079363d31e087aabe24b345588f5996a665f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7a55658b2d19b9e8311dc79a5e3d51fa1296c6a17ce250d083ecea1f92a5bc3d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c328bfa734b290f1b4eb9308df7b36e60d09221fcfd313d5b8dff8254bc121a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989825], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "4f952a2a-90c1-58c6-8689-895422b92d70", "validation": {"axes": {"B": 1, "N4": 391, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.007083912845700979, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4f952a2a-90c1-58c6-8689-895422b92d70"}}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/14f79d894fa215c9339a45e31034cf053b0394262718ea5b2ef6f73b54ccbce8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/23b375470dc83c854e854da29dbd37b210429f265dabb2617a8d8dfd749df6d0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ec30ba59e573958d9b848368683f7db1de30e14fa286d535215816e699cab27a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9064b5aa0898f9e54372f27d8f13a9da5713240e2a287800ae29bd33234a27af.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989826], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "7e4e5d7e-10bb-58f3-ab0a-0053dc4a910d", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018532858230173588, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7e4e5d7e-10bb-58f3-ab0a-0053dc4a910d"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/c66df53e4e2aa9a78dc3478c1670b93772c7fcd09e9bc86a5b8dff72dfbb4b67.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989826], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "1034f7bf-069f-53a4-8a07-bac0fbdc3eec", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5375021840445697, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1034f7bf-069f-53a4-8a07-bac0fbdc3eec"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/267eaebdb6dc0081b729b82939fd95b36309a3731ed529e30205aca5f3117c28.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989826], "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "926dc757-1514-5c87-9773-99d219d9f473", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.30510560562834144, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "926dc757-1514-5c87-9773-99d219d9f473"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 440, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/ff5eccd42de498fb60561c7370b15fdca4468368b975b6d7ec87ef54cca114a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/0f4aa9e73bbb46845c3094cef2c38793f32505cb1e26f4a01b8c33863e82ba6f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c328bfa734b290f1b4eb9308df7b36e60d09221fcfd313d5b8dff8254bc121a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989826], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "35baf826-1c46-56db-a032-0e97cc76924a", "validation": {"axes": {"B": 1, "N4": 440, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.005181238986551762, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "35baf826-1c46-56db-a032-0e97cc76924a"}}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7751995cb5b9fbf8fafe4ec7fb51da0fb18c1d1cfd2c9faa075bb48eaf3a994b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/5dd928e2d9e85109b53320ff46bb9c5c9700121862c43b46d506d001fe64ca04.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7ac0d26e282cd34afaf272ae7ae512f07c41ffcdcc57710130808c7e5b883361.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8afd77630b6e854adf4567af14d3a5b00baa014baddf89b587821ca0cd895cef.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989827], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "22590027-953c-5b6a-b2ca-508afff0b457", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018935003317892551, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "22590027-953c-5b6a-b2ca-508afff0b457"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/8d8ca18ba0307591bae6d6fc667e8672dbee0defe7d930121606b0ab8c3ead98.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989827], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "5dcaee1a-f718-5532-9191-3024e0c3057c", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5372484093531966, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5dcaee1a-f718-5532-9191-3024e0c3057c"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/b4c92f2ff6792f76b3d3cf6dec8d14c2f8a488e0834f9642a24f0f6b67e616be.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989827], "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "6aff8387-0383-52da-8d07-d0c10ebf2d4d", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.30552644236013293, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6aff8387-0383-52da-8d07-d0c10ebf2d4d"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 447, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f95d9ed17c639402c86d27fa8c6689ae0c77953699d42251e139324648a09f9b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/419ba765f9c0892cd0f84aa2f72e8bf26bc0faf25efb28807da16da3cd0ff135.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c328bfa734b290f1b4eb9308df7b36e60d09221fcfd313d5b8dff8254bc121a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989827], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "f9f27f44-c195-594c-ba33-277c731bc996", "validation": {"axes": {"B": 1, "N4": 447, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004752210807055235, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f9f27f44-c195-594c-ba33-277c731bc996"}}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9c2d6fb4390c45690358c829016c5b70eb2f2ed17c3ab281a0a098a0e2726ee9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b0d280c7576b9f954e8fcd448c036b24e6450ee531d155ede8e9e06f364e5e90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/0524c704a60b543ffa7575912a6dbcfd09bd0c50dc80e474a7987514bc45e0dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/1c6ce462fa6d7422ae594bd5e796548deeac649395162b0fb6154976a6613c98.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989828], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "6dd20c52-b8f7-5c0a-9a19-40c3e9f5c8d1", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0018482441082596779, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6dd20c52-b8f7-5c0a-9a19-40c3e9f5c8d1"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/0630a5b8fe5b05b1e966f162b9af9ad4deab43a79637a4e352db98917c34bf20.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989828], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "5e05cdb6-af1d-5a5b-8be8-0341feb4e34a", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5389163577929139, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5e05cdb6-af1d-5a5b-8be8-0341feb4e34a"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/88b3cfc2aaa50bb57cd85307bffeb877a6964104c5dfb13e7104162f68cd43c0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989828], "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "3e554d41-bfbe-56cc-a05b-179acb8a60f4", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [7.450580596923828e-09], "seconds": 0.3053081319667399, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3e554d41-bfbe-56cc-a05b-179acb8a60f4"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 402, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c96a79c7877a3402cd92e63fff2c5814e620805d17eced34aae1d97dd42a2bbf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7e43282fa8e9a369fae67d91337a6a5e2a8b4d4bab2cbbf856055da844dff505.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c328bfa734b290f1b4eb9308df7b36e60d09221fcfd313d5b8dff8254bc121a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [989828], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "a13e6ed2-56fb-5d60-97ae-1abffcfe4ce1", "validation": {"axes": {"B": 1, "N4": 402, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.004425243008881807, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a13e6ed2-56fb-5d60-97ae-1abffcfe4ce1"}}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9363d93c51932542601955d0e76ea3af6382b6534d3576e5e2d4db8abf149d53.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/e032ffbbf4230e554361cb952e5006e5718589bf65ff403cbdbdb71b210b1992.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7b40b0acfab32497253315212a3dd11cfe314eb7f53b26bd27b516dd8102c470.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/87f2b7e6239721531e6b775ffaed1c9dfdf3c9102020db0495a11033fcbddac2.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [990426], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 601, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 601, "uuid": "1b7674e8-c16e-54fc-98be-3fc37691c059", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.0018679341301321983, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1b7674e8-c16e-54fc-98be-3fc37691c059"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/f9e230d898ba149d5c73c09a3ead21604daa0f69eb4071799ae532c6c4b33481.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [990426], "selection_step": 601, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 601, "uuid": "40955dc0-c743-56ea-9809-1954d1c880ca", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.6982219400815666, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "40955dc0-c743-56ea-9809-1954d1c880ca"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/af2c27bbdb367588002852e0c04e855500077bc9b5f6e5a7080fb1e8bbf399e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [990426], "selection_step": 601, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 601, "uuid": "7f2b006d-0a89-5e93-abd0-f8563c06f8c9", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3068130211904645, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7f2b006d-0a89-5e93-abd0-f8563c06f8c9"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 370, "N8": 1, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/a345afa325b0fb8176f03830dc1c14413bee3ecbec38e459b824180d90295ca1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/89723945dfacec4354824e18c80b1873888eafd5b50ec9334cc1dee82d76b1fb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c328bfa734b290f1b4eb9308df7b36e60d09221fcfd313d5b8dff8254bc121a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [990426], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 601, "uuid": "60d85ebd-5bc6-5fb7-8062-e4646777caea", "validation": {"axes": {"B": 1, "N4": 370, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004576727747917175, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "60d85ebd-5bc6-5fb7-8062-e4646777caea"}}
{"actual_requested_output_tokens": 1204, "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/09992723f59f11433609b144a3ec27b3fbc094e7f6fdc520693b5b679ca1a403.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a6854895c6fef58815a8941c4594e1dee46dd79f135be4c0b1c4a44f797a497d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/473bba7a3533a50f39cb0c551f5c0752f583d0d17df2fe891d0c684c951fce3f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8782c87d4f937417902687fa53270c94a20a8e34ef32a5185afcd51dfe829eaa.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "layer_id": 0, "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [991027], "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1202, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1202, "uuid": "0cf25067-8587-5fe2-b8b8-3093a243e8a2", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.4901161193847656e-08], "seconds": 0.0018994337879121304, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0cf25067-8587-5fe2-b8b8-3093a243e8a2"}, "variant": "decode_1"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/0dc6aaa1cf33dfffbdedbb27a96f472dc97d2da50a90cab5c2a183ea540130e9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [991027], "selection_step": 1202, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1202, "uuid": "888761a7-d8c0-5d0a-9e2d-e206473eb438", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8692195420153439, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "888761a7-d8c0-5d0a-9e2d-e206473eb438"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/1136ae48ecf37fd4fff834667dd5024df6acab6ce12040204301159f20399590.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [991027], "selection_step": 1202, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1202, "uuid": "0797a69d-b39b-590d-a7bc-2e7896a70793", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.3060846603475511, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0797a69d-b39b-590d-a7bc-2e7896a70793"}, "variant": "decode_1_cached_False"}
{"actual_requested_output_tokens": 1204, "adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_profile": {"block_size": 64, "branch_id": null, "hash_id_scope": "local", "in_256k_variant": false, "input_tokens": 989824, "label": "full_context_max", "locator": "requests[1920]", "model": "claude-opus-4-8", "output_tokens": 1204, "reusable_prefix_tokens": 987584, "reusable_prefix_tokens_256k": 253440, "role": "main", "selection": {"field": "input_tokens", "note": "Near-million-token tail; stress case, not a typical request.", "percentile": 100, "target": 989824}, "session_id": "5b6e131844fbcb4a57f8cf516831dc679004", "t": 192785.654, "uncached_tokens": 2240, "uncached_tokens_256k": 736384}, "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 416, "N8": 2, "P4": 128, "P8": 256}, "cache_predecessor_locator": "requests[1919]", "capture_decode_steps": [0, 1, 2, 3, 601, 1202], "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/68b09500e07c66f35fa9eab66df0eb7d88296a2f9b71cd5a01aac064a7766d15.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "input_ids_sha256": "28f0ba8d30cb3d2adef488d08b0f1e7952d8f2f679a3eb3355ba465ab9fcf012", "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c2cf5c31d7bb98642318c0acfc69a7b8d72a60bb0bf93e8e1a3010d893b4c97b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/bac456d3b5829f27b9fa8fc4259ddb68d857def1e4ec3e030065043e5491b0bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_full_context_max", "scheduler_mode": "decode", "scheduler_sequence_lengths": [991027], "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1202, "uuid": "e40b517a-47c2-5f8d-bb80-53962880c9e7", "validation": {"axes": {"B": 1, "N4": 416, "N8": 2, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.005160517990589142, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e40b517a-47c2-5f8d-bb80-53962880c9e7"}}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 4096}, "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/764855204f9cc3e6f07d91ad9d1c08d07bd9fbf315d2e35a55ea5face56b9e78.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [4096], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "5bb71426-9b5d-5301-a44d-b6b4afc5a924", "validation": {"axes": {"M": 4096, "Nkv": 4096}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.27322842460125685, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5bb71426-9b5d-5301-a44d-b6b4afc5a924"}, "variant": "rows_4096_compression_0_cached_False"}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/cfa6bbd7d01286eb7e5d7c65a7008d726d7973153a4757b10425033eeaa7292e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8f583f17e0de173de6c6c4efb3d1843a5d9e14e2f1e733b1184b0eecd1fc5eb8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/c9fd4bfc697b50f85dfa06eabe24a02264bd310c0e0ae3b53525343a45cb663f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/66c3deaf8590c8e10a82c739636202fc7c4666ab9abf7ac655a24f39a3a6911b.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [4096], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "1a7483d5-5abb-5604-9217-2050ffde435c", "validation": {"axes": {"M": 4096}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 3.2782554626464844e-06], "seconds": 0.009584414772689342, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a7483d5-5abb-5604-9217-2050ffde435c"}, "variant": "prefill_4096"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/78724af1bed4afa2eccf9a9528b1bf4a13639f7ccb76d71b997a64e8d0f88373.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [4096], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "0b1815f1-1059-590a-87ae-21879f49482d", "validation": {"axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 1.2033551116473973, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0b1815f1-1059-590a-87ae-21879f49482d"}, "variant": "prefill_4096_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/f93666ace0ac315f3e7f7c898091c58a43abab7dd7f0447c99545c13e6b8ac56.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [4096], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "b878a86b-c444-58fd-88fb-935d7afcaff4", "validation": {"axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.4996452019549906, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b878a86b-c444-58fd-88fb-935d7afcaff4"}, "variant": "prefill_4096_cached_False"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 6144}, "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/d5c32961f1ec9ab5d15516cf5c24897dc1c20ddc1dd2e65dd77594433dc1dcbd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [4096], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "b57f47c5-314a-57db-9233-9b1470763ce5", "validation": {"axes": {"M": 4096, "Nkv": 6144}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.21970500983297825, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b57f47c5-314a-57db-9233-9b1470763ce5"}, "variant": "rows_4096_compression_2_cached_False"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 8192}, "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/99c0d5885a99174b4a3b0c4c9da7b57fbdbd07bde57786f8a39e9738972edec6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [4096], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "7fd83f55-ee1c-5bec-b886-047096621576", "validation": {"axes": {"M": 4096, "Nkv": 8192}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.1358597120270133, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7fd83f55-ee1c-5bec-b886-047096621576"}, "variant": "rows_4096_compression_1_cached_False"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 4223}, "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/55cd5d99e9988e623c504a4e39860ab107cc895b5189159b0adcc3e4a5665dba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [8192], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "f475abca-84d5-58f9-a8b0-ee05530a4fcc", "validation": {"axes": {"M": 4096, "Nkv": 4223}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.22201813105493784, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f475abca-84d5-58f9-a8b0-ee05530a4fcc"}, "variant": "rows_4096_compression_0_cached_True"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/9e57f0d9bfa3a180ad9c23f084e909258244c223d642e216ff9e3e64118fd7dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [8192], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "75452890-0209-5264-bb51-e75d6c5b75c0", "validation": {"axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 1.190809722058475, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "75452890-0209-5264-bb51-e75d6c5b75c0"}, "variant": "prefill_4096_cached_True"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/18b0875290346a2ab8cf87edf2c7e79f3a74ed66d74dd0e5e3dea0e5d69096e0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [8192], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "b1fb4669-1248-56ff-978d-f610aa47ea2a", "validation": {"axes": {"M": 24576}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.4897424881346524, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b1fb4669-1248-56ff-978d-f610aa47ea2a"}, "variant": "prefill_4096_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 8319}, "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/e2f6fa28838a8933d169d786f4d68604a5e325245b90bb1c9c32ffb122ef08ac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [8192], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "f0c8462c-5541-51c8-90c2-52686dd93402", "validation": {"axes": {"M": 4096, "Nkv": 8319}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.13585394481197, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f0c8462c-5541-51c8-90c2-52686dd93402"}, "variant": "rows_4096_compression_2_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4096, "Nkv": 12415}, "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/7ec3de10ef7e2994921ef173190d2e76ded5926feace029a3f002aafdaf17ecc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [8192], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "b1ee31d7-328c-59af-a60c-189c90e2d248", "validation": {"axes": {"M": 4096, "Nkv": 12415}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.13608363270759583, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b1ee31d7-328c-59af-a60c-189c90e2d248"}, "variant": "rows_4096_compression_1_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4032, "Nkv": 4286}, "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/a77bf95a04fd1b3eb9e91394d5968fd63fd558730f3317a6c98ec43eafdc0ed1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35008, 34560], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 8, "uuid": "c6c6e0c9-61e2-505e-831a-fef983c9dad9", "validation": {"axes": {"M": 4032, "Nkv": 4286}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.21675625583156943, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6c6e0c9-61e2-505e-831a-fef983c9dad9"}, "variant": "rows_4032_compression_0_cached_True"}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4032}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9fa83b55031c6da0f4dd6eb39e2090c86cf8c347ebc6ec6371a5d9b6a6c107a4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/73ad4498b4d4e60062bac30e73443aca13ad6669594a01a10173b1b70fff8f61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b1569af7a8e420dc9b2db952010a7227f25877371887f86c373026eb720bc9b4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8f322960affe94c26245d79f46be5dfae183c3769ac3c07c935d3dbb3d345ed3.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35008, 34560], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 8, "uuid": "ecdd3d80-02ae-533f-a687-198443ccfdca", "validation": {"axes": {"M": 4032}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 3.5762786865234375e-06], "seconds": 0.0119220782071352, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ecdd3d80-02ae-533f-a687-198443ccfdca"}, "variant": "prefill_4032"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24192}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/45ec6ab89391b2b94ea6f0c220b32849c43f721bca937cbc6936ea8f17271ab7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35008, 34560], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 8, "uuid": "c24c30ba-f971-5480-9756-4906df92201e", "validation": {"axes": {"M": 24192}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 1.2006504167802632, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c24c30ba-f971-5480-9756-4906df92201e"}, "variant": "prefill_4032_cached_True"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 24192}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/04885efb8cd8095b0626dc86ce33245cf9403d3c75587cada67df74ecf0116f5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35008, 34560], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 8, "uuid": "a51a750b-1fee-5798-ac8c-6254b2d4d4f8", "validation": {"axes": {"M": 24192}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.48242560867220163, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a51a750b-1fee-5798-ac8c-6254b2d4d4f8"}, "variant": "prefill_4032_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4032, "Nkv": 39070}, "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/2099ef68a72e44bebc6a4ea7e0f0001d2322a16834423a65594d9a445e44f9f8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35008, 34560], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 8, "uuid": "2c6e4648-5da2-580e-a26e-46e6f300b0a5", "validation": {"axes": {"M": 4032, "Nkv": 39070}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.14000864699482918, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2c6e4648-5da2-580e-a26e-46e6f300b0a5"}, "variant": "rows_4032_compression_2_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 4032, "Nkv": 73854}, "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/c225935ec58ab02c9d3e34267901d49661d41fbeff521fb7c95756e851cecb5c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35008, 34560], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 8, "uuid": "6385d1ff-f172-5c14-b73d-1c2b5943f045", "validation": {"axes": {"M": 4032, "Nkv": 73854}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.14596810471266508, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6385d1ff-f172-5c14-b73d-1c2b5943f045"}, "variant": "rows_4032_compression_1_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1984, "Nkv": 2238}, "compression_ratio": 0, "definition": "sparse_mla_prefill_h64_d512_k128", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k128/2d74e9a5b97064d349ebec93d921024706e89a2912007c7808e0af74652491c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35392, 35456], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 9, "uuid": "aa1f4b6e-c84d-5723-8007-090ffa76004f", "validation": {"axes": {"M": 1984, "Nkv": 2238}, "definition": "sparse_mla_prefill_h64_d512_k128", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.12401726329699159, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aa1f4b6e-c84d-5723-8007-090ffa76004f"}, "variant": "rows_1984_compression_0_cached_True"}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1984}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b634502a4eb52d5e6f5578329bb2d694f92bf423d0b47d5d33e80542b8a357c9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/267ec2804c72a33a306ad7c882ff27c7f688b188fe43c392d4f1efd369d694d0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/3b1bd47e74b20959e285c20b5ac47a292018aef09eabad9bed688bc50f80e396.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ec13eb5bcf3dba7f9855ce0baf02107ea76c9335237b05142b4e559a5619e419.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35392, 35456], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 9, "uuid": "491c4277-463b-57a8-b65a-38cc0439366b", "validation": {"axes": {"M": 1984}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 3.2186508178710938e-06], "seconds": 0.0095808538608253, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "491c4277-463b-57a8-b65a-38cc0439366b"}, "variant": "prefill_1984"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 11904}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/76f5455e6821e3a552ca4714431200a1cc59f16e82526871fe6226e1fa6090dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35392, 35456], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 9, "uuid": "7e7da669-03fe-5ee3-a1cf-b496e7614a56", "validation": {"axes": {"M": 11904}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.8494550948962569, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7e7da669-03fe-5ee3-a1cf-b496e7614a56"}, "variant": "prefill_1984_cached_True"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 11904}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/82ec663da9cb8e1785b6e27599ee8abad1158d6200ae714513bfe17c3d55d101.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35392, 35456], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 9, "uuid": "7db1bb57-9034-57c0-95b3-2a45bafb1409", "validation": {"axes": {"M": 11904}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.45867891516536474, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7db1bb57-9034-57c0-95b3-2a45bafb1409"}, "variant": "prefill_1984_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1984, "Nkv": 37662}, "compression_ratio": 2, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/593b16763e612bc299b370621427e32bd9fc285a2e74bf1e3223cde14f51c7ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35392, 35456], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 9, "uuid": "0b988a7a-2b52-5f49-8c99-eaa373baf797", "validation": {"axes": {"M": 1984, "Nkv": 37662}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.08471962995827198, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0b988a7a-2b52-5f49-8c99-eaa373baf797"}, "variant": "rows_1984_compression_2_cached_True"}
{"adapter": "Gather all logical query heads and outputs from tensor-parallel ranks; retain the complete BF16 workspace and original sparse indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1984, "Nkv": 73086}, "compression_ratio": 1, "definition": "sparse_mla_prefill_h64_d512_k640", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/sparse_mla_prefill_h64_d512_k640/719f6f15791967aa4272baf8c5bf3c461f013c11f949a9d16eb478250befd9eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "sgl_kernel.flash_mla.flash_mla_sparse_fwd", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "prefill", "scheduler_sequence_lengths": [35392, 35456], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 9, "uuid": "43e73df3-0172-583e-944e-789dff99b8b6", "validation": {"axes": {"M": 1984, "Nkv": 73086}, "definition": "sparse_mla_prefill_h64_d512_k640", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0831243572756648, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "43e73df3-0172-583e-944e-789dff99b8b6"}, "variant": "rows_1984_compression_1_cached_True"}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2d668b2bdde0e5a33414874ad0837a906621f9e319fb4ee5a6ea4f6cdfc4711a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ad002998d952e2da467a96a7f9512e4426c249e0ffec6764bf79aac23d572001.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/70607226c53c2d7251994f07fd4ea48fce00b8309715f0d3f68d5bef1de7a86d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/a446ce78deef3c4deb3c519281a19b6fb7897115d9612d2644dfb155992c36fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35009, 35393, 35457], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "8fb5668d-9ace-5884-ae82-5cfb3e534ef0", "validation": {"axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 4.470348358154297e-08], "seconds": 0.003166140988469124, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8fb5668d-9ace-5884-ae82-5cfb3e534ef0"}, "variant": "decode_3"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/9d4842fc59ce46e337942b1cba771d037ff740095aa7142979d49b7aae1029ca.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35009, 35393, 35457], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "b4fecb00-3198-5a46-b718-a14bc4361694", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [3.0517578125e-05], "seconds": 0.5468553341925144, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b4fecb00-3198-5a46-b718-a14bc4361694"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/c9a161ba440d995c1fef182947f196952c92f8a1526074b4c4496a7836f1ee5a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35009, 35393, 35457], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 0, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "69b2486f-f566-5982-ac2e-8c800ec1dee5", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [1.4901161193847656e-08], "seconds": 0.31446270225569606, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "69b2486f-f566-5982-ac2e-8c800ec1dee5"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 3, "N4": 146, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f5df05e0ec51a60f36ca4e451427d94c1f29a31b08b3c08f194600ceb579b7f3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/a2896a2c7a77e677a03469000528f09853cb68ca8f38829a16dbe1df13ca2527.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c9f52977c1c53b43c87c856fd4e33082f009452798662f3300c78b31a1ae9245.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35009, 35393, 35457], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "506de298-8285-52bb-87c1-15d03a7a50d1", "validation": {"axes": {"B": 3, "N4": 146, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.006262518931180239, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "506de298-8285-52bb-87c1-15d03a7a50d1"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/39a8b50e8b4fc730544481896881e88b41a921174982aae68b63bd2f1b6ef0c0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9a4526719c278596443909500b38e29f7d50bbc8cd648b825b9e783f57313aff.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/15e101bc4551272a345ef15e7c05f2bbb66e2705bfc58c2be631ad5245f8256b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/abc5eb2e8216456dfcad63b2ee8ddc72080be9456a0539f8856ff68ef505a0ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35010, 35394, 35458], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "b22e0ab8-07f2-5028-b601-d5e2771d3d2c", "validation": {"axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.003138778731226921, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b22e0ab8-07f2-5028-b601-d5e2771d3d2c"}, "variant": "decode_3"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/0ba220fbeebfbd704489d0a0787e0cd8693745eae6b39d0c2ac80f628708bd38.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35010, 35394, 35458], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "d4b15562-fea6-5caf-ba58-a6ede4195c3f", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.001953125], "seconds": 0.5430445061065257, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d4b15562-fea6-5caf-ba58-a6ede4195c3f"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/e27eb4ceb7a5548020a18cb0a1588d8b56f2419bdabdea8cf58afa6fa22cb0cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35010, 35394, 35458], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 1, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "22791a4a-6151-5445-806d-c3793c972208", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.00048828125], "seconds": 0.31428917311131954, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "22791a4a-6151-5445-806d-c3793c972208"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 3, "N4": 147, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/82be75d76af79c535b3246ec7a5b9266b1797e6d83adf6b67b6ab143f136b947.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/8dac3a9477bb14818a835bc00f9c8d7e25000c1808760c2a1c989d690bba8818.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c9f52977c1c53b43c87c856fd4e33082f009452798662f3300c78b31a1ae9245.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35010, 35394, 35458], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "280fce72-38f0-5ddd-9ce7-ed847babecfc", "validation": {"axes": {"B": 3, "N4": 147, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004787764046341181, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "280fce72-38f0-5ddd-9ce7-ed847babecfc"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/80585fbbd2d8cdbf3deccb7fee2ca53482729bccac0be1297ffb561a0e93f73b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f8674e58b6644b08d75d019456aad2d848727ba7ee540889e181818e7139451f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/eb188f2e6a71a78009018322d159f5cc600e84faac6672bc00c9f8fd0b8679ea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/19405bc4c99b9d5cb7b352b177a855747c7d6a7ca93f9d29324076ec696ed8da.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35011, 35395, 35459], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "c25f65ed-4124-5afa-b47d-07d571b319eb", "validation": {"axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.015995109919458628, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c25f65ed-4124-5afa-b47d-07d571b319eb"}, "variant": "decode_3"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/7341d2ea5331fa1ffccd2b88437fa8179fa288eed5331e0053d1e4b20132799b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35011, 35395, 35459], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "3dfd91e1-59a0-5a61-89b3-3434003ad831", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5427321679890156, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3dfd91e1-59a0-5a61-89b3-3434003ad831"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/c0c963830f7d18156b7a27526c487f0d00a1dd6d146cfd57b05a18a61b048e99.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35011, 35395, 35459], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 2, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "6d903c42-3165-5ccc-a2c3-3ddfc8a7ed4a", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.47860030410811305, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6d903c42-3165-5ccc-a2c3-3ddfc8a7ed4a"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 3, "N4": 145, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/cebc0f71b12521cca5bf67fce10baeba1bcc98fd2e1b27133a688dfde116b3f8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2f19143111599a342f90237ebbbec640fd7ee106a6c7c37c90d1f21bb3b67dab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c9f52977c1c53b43c87c856fd4e33082f009452798662f3300c78b31a1ae9245.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35011, 35395, 35459], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "9e1eb156-12ab-5687-b99e-5694bc8e11f9", "validation": {"axes": {"B": 3, "N4": 145, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004719646647572517, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9e1eb156-12ab-5687-b99e-5694bc8e11f9"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/9cf39b82297378f66b74bb51c14c116a4fe474b8121cfec3cad4e2e39f6d94ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/c4087880fc82c8f309977295103024aa2f291cd4c23df6d821ed8ccfa924397b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/5ced3d55cc98d8df048993f5dba53f0886e71859460bbcb48bc7c64356e76f09.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/5bfdc7caeb5ed03f6bb269e131ee22b3bda2320742229e6eb1c94de77b5b823e.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35012, 35396, 35460], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "bf11c6be-1cf2-5593-8167-3d13642ce599", "validation": {"axes": {"M": 3}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0033489898778498173, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bf11c6be-1cf2-5593-8167-3d13642ce599"}, "variant": "decode_3"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/bc843cbb7c7af673836e7f79d508376468f318df69759e78653a5e3dbc6d0165.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35012, 35396, 35460], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "759f9ddc-17f5-5cdf-b9e6-d7f43e5ff965", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5432731551118195, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "759f9ddc-17f5-5cdf-b9e6-d7f43e5ff965"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/05cd5b433dcc242dfc03139f3eba7e832eb23d5289fc4770b0b4435c6ea89fc2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35012, 35396, 35460], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 3, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "82fbf85f-f1f5-58fe-a053-583a68c3717d", "validation": {"axes": {"M": 18}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5164944920688868, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "82fbf85f-f1f5-58fe-a053-583a68c3717d"}, "variant": "decode_3_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 3, "N4": 145, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/f1a17d4804206a6ca6fffe747dedcd6031e3760933e0eea8f6515c0446bf974e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2f19143111599a342f90237ebbbec640fd7ee106a6c7c37c90d1f21bb3b67dab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c9f52977c1c53b43c87c856fd4e33082f009452798662f3300c78b31a1ae9245.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35012, 35396, 35460], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "c4bb3ecc-abd1-544d-933e-5dfe11ca7428", "validation": {"axes": {"B": 3, "N4": 145, "N8": 4, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004712217953056097, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c4bb3ecc-abd1-544d-933e-5dfe11ca7428"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/4514ca9b389e49bdc0ec882f4d49de1773fdbd71eb196a0194873eff48846acb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/3e47a783f52602828428b6c3aa646e733a4cf5c3bb78f9a6d430178c59f24c96.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/38b1fbf2d0d656032280600042baca8806ddb783b11860ae3ff6c6bb964aa553.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/92a9d60c184a26b04fd4bba0ce7c80ef32c997741527ff1e8870b7003135d185.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35116, 35564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 107, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 107, "uuid": "fa53421f-e2db-580b-8e24-ff234a18d7a9", "validation": {"axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.003359489608556032, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fa53421f-e2db-580b-8e24-ff234a18d7a9"}, "variant": "decode_2"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/9aa9f5547e4f207fc0d859025d8779b2fc61cc3ae8bc068c09b6858f547fe3a0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35116, 35564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 107, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 107, "uuid": "bd5539cb-9883-5d6b-b935-0fc2eabb4551", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5384206641465425, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd5539cb-9883-5d6b-b935-0fc2eabb4551"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/2d3b34d5fa8f1e8653b055649159a24526cea9acb5b8e1a471b288e1cedbfb10.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35116, 35564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 107, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 107, "uuid": "9f6a3944-09d7-52ca-abb8-4b01960f3577", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5091732949949801, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9f6a3944-09d7-52ca-abb8-4b01960f3577"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 2, "N4": 135, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/9b98d9134538d7ed172ca951a5997176074af8f11e26b355a8ad8fd9b8b02f51.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/e9f99b6f57813b574a27f40aeda2b0ea6182df8eacf37cc74abcf8300b67e33d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7175e91a4f69f8fbae97e298a50b1e0fdd3c8cb42121c0b304f05a4db79f33b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35116, 35564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 107, "uuid": "193a3358-bd3f-5f02-bd91-33e45717225e", "validation": {"axes": {"B": 2, "N4": 135, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0038592899218201637, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "193a3358-bd3f-5f02-bd91-33e45717225e"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b0931ec9a31a05d20e3024ffafabeefbac295d62eee347ec95ee3f9fe57a1256.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/90f02c869fa8f5456ca9ed87f62cc4e755212f1b0126fa59c90c56475a2af94f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/81277ea0c24f35dac23469d46496c069612cbb669ff0f3c9deda99c6804be28d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/ad64fa161fa575df2e029246f649b0e467687820c41e36f4f32c1a535ead015d.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35117, 35565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 108, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 108, "uuid": "82dfd151-92c5-5eff-8342-19b4e342569f", "validation": {"axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 8.940696716308594e-08], "seconds": 0.0030926777981221676, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "82dfd151-92c5-5eff-8342-19b4e342569f"}, "variant": "decode_2"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/25ca1d7e28d2c6cba057dffde7fe560203d0e5120467cf76a5cf361988c294df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35117, 35565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 108, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 108, "uuid": "d500197c-537f-5a7d-b132-5dc3b4651d12", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [9.5367431640625e-07], "seconds": 0.5360788740217686, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d500197c-537f-5a7d-b132-5dc3b4651d12"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/c90a693e662936f2e4593b1e08a01dc9726cc6751aaee18a53b47201a98d8b55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35117, 35565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 108, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 108, "uuid": "f2fe6eb3-e86c-5e2f-beb3-1de5680a8044", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5105185238644481, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f2fe6eb3-e86c-5e2f-beb3-1de5680a8044"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 2, "N4": 142, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/eef89e0c51aee47176fd9da00ba46ee659fab25e919b56d9e7d99809164e0196.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/3ab1b2be5aa477011182d95d4983647dbf790bf20f53898ab38e964b9b7c1f33.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7175e91a4f69f8fbae97e298a50b1e0fdd3c8cb42121c0b304f05a4db79f33b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35117, 35565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 108, "uuid": "12180b5b-0be9-5de0-a091-1bb8e20af73f", "validation": {"axes": {"B": 2, "N4": 142, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003872013185173273, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "12180b5b-0be9-5de0-a091-1bb8e20af73f"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/472c45c2f00070e7566940168061d24f47d66eda3f7779331289ad933fdf4a88.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7356a566ff0dff2ee6fd902b3fb59220abc2de2ef3e699e4377e9e0bfa3d8d23.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/65163185b87a691c8bcbb1c1f6601326f8bb79d7a5bd27302e3cf54e7ee87c3c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/fc0bb4c3b4d5eb744961c8185b55d4f956d89f53d3496afa8fa73bb596a3541b.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35118, 35566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 109, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "4082bf7f-299b-5d20-bf12-133e52e792ab", "validation": {"axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.003034201916307211, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4082bf7f-299b-5d20-bf12-133e52e792ab"}, "variant": "decode_2"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/564affa818afc21e5f7b864eca737d89e1a78824f63d9d08feb4256f3133a49d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35118, 35566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 109, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "fc247ae3-7332-52cb-9e37-11d6417daf81", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [1.4901161193847656e-08], "seconds": 0.53823347017169, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fc247ae3-7332-52cb-9e37-11d6417daf81"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/98e010945f80933ba933dd6d90d79797886596a03e6f50ec17a7b1bf970b13dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35118, 35566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 109, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "1a62ee81-50f8-526e-a53b-f75d3f5a03d5", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5099127921275795, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a62ee81-50f8-526e-a53b-f75d3f5a03d5"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 2, "N4": 142, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c2718b594ea5f338f2a08db9a8fa70031fdafddd79b79cb6688cc0149113e78e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/2b2069cb79c87ef945fff06ec4c6e5bd5820f1153b05f4616f13a73a371ba410.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7175e91a4f69f8fbae97e298a50b1e0fdd3c8cb42121c0b304f05a4db79f33b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35118, 35566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "35704cb4-9369-54f7-9452-896d00e9f9d6", "validation": {"axes": {"B": 2, "N4": 142, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003853880800306797, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "35704cb4-9369-54f7-9452-896d00e9f9d6"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/b75c48785f6884ee69a0d69e907e741bcfa70ae9aead73ef7e5391dc5765dcc0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/290a3f5b4e61b941e132d9143506f5fbf714be34318acf88058c19b294873f60.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6ef1b7fae7435e3175209b0c9cd75de8b4d862f3054d3ca151763fc0a88b8d4c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6fe15335ee2427356d678dea6d0d4530917735f05ecf27c4aa31df966f8e99b1.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35119, 35567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 110, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 110, "uuid": "99713957-37e9-559a-92d5-a800373b3e75", "validation": {"axes": {"M": 2}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0032311580143868923, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "99713957-37e9-559a-92d5-a800373b3e75"}, "variant": "decode_2"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/7171098400c11f313a1f99a27d1b25d64ee3b9bf0e862472f9b6de4d35eb9d26.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35119, 35567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 110, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 110, "uuid": "54cc9aa4-83ff-56e0-830b-a9634ca9a54c", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0001220703125], "seconds": 0.5387282078154385, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "54cc9aa4-83ff-56e0-830b-a9634ca9a54c"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/289bcc3be71bdaf8abb4dd561dbdb7612c4b546a483e6c83116a4e6ca6f2ae71.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35119, 35567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 110, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 110, "uuid": "0c7231d6-7f41-572a-ac47-eceb974cad41", "validation": {"axes": {"M": 12}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5106046716682613, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0c7231d6-7f41-572a-ac47-eceb974cad41"}, "variant": "decode_2_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 2, "N4": 135, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/b396ef2bc6965cfd6bd24b241b84110341cf817221e89c196d9276ed9636b489.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/bc1633e87b1e444fc238af8089f6989762b95d63a2fa9fefae3203ca1035f1a0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7175e91a4f69f8fbae97e298a50b1e0fdd3c8cb42121c0b304f05a4db79f33b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35119, 35567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 110, "uuid": "b74cdd36-78e7-5612-ad84-cc44b15c5b29", "validation": {"axes": {"B": 2, "N4": 135, "N8": 3, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.00390625], "seconds": 0.00393477501347661, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b74cdd36-78e7-5612-ad84-cc44b15c5b29"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/380c748ec6377c69e4f4dae5bdd8ad09cdb1446b65483d1ac54eaa5ed9c20963.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/0b46ff7dcbf13b37156459f61cfc5a43e0c825f0abaf8aa98f860fb2c04abba8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/1c2b58efbd07006dd8346ed6c32c17893192685781f7c5f84dc27f47032acf8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/2fae900b61c5c2f6e097d6a9efb4e2b8d3545cf6646138e166d9e3ec930ac853.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 119, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 119, "uuid": "eea430f9-5bcb-58a9-b079-c1bd8f062f3b", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 1.1920928955078125e-07], "seconds": 0.003239080309867859, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "eea430f9-5bcb-58a9-b079-c1bd8f062f3b"}, "variant": "decode_1"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/0aa68cce6b21285e516bb732cbbfd2fb48834a4d2a70dcddc159d1cb8039024a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 119, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 119, "uuid": "6e152962-f1bf-5b6e-aaa2-e0a40cebc869", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5724243517033756, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6e152962-f1bf-5b6e-aaa2-e0a40cebc869"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/225488cfaa707501dc60747f67d8baefe00d1536d12bf0cb329c999018ab691f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 119, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 119, "uuid": "925a836d-0688-53f3-bcd8-c8b1a67135dc", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5190592082217336, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "925a836d-0688-53f3-bcd8-c8b1a67135dc"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 122, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/79ec5b681bdaf2575efced183ff5b8fb657722eaab94c4e6080a6bbf12d729c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/e0c9aeb5080f443ef65a4ade56b5d1ff3007998177e9468237fabde133ec4755.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d2fa94370cd62981deb374c2e9d7dff2abdb59ac751a6da4a55b2514f87fd6de.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 119, "uuid": "0f727df7-b91d-5b0e-a1d7-0a24dc243af6", "validation": {"axes": {"B": 1, "N4": 122, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.002870344091206789, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0f727df7-b91d-5b0e-a1d7-0a24dc243af6"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f3e16897e5635b06ad236a9b0159273b54386b9d2d4b728ddb5fbf31d40f7893.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/7609f7c926b42d80cc4f7a2b942bdafd6b7a4d932561756a9a11fbab37c821ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/c5fc2c52ac0c28f18030c03f884ca300f1632e20208d62c300200af562b711c9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/f0788eae3193a3437bd228993cf5e0cf56c91cd7bec90c007f333b42a6f2c02c.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 120, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 120, "uuid": "51a3452f-7f5d-5c24-80a6-8c2f7f9bcf23", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 2.9802322387695312e-08], "seconds": 0.003138733096420765, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "51a3452f-7f5d-5c24-80a6-8c2f7f9bcf23"}, "variant": "decode_1"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/9f956d858cf6f80a86747eb7a5795eec2a713e598036d9ae80d4c96c8c119983.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 120, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 120, "uuid": "e8f39e29-79a8-5f9a-9e6e-931a67868ed8", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8617571918293834, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e8f39e29-79a8-5f9a-9e6e-931a67868ed8"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/0871c01e37b7892d7b53ef86b9c2d10365f801a77d2922b2a7f5c57a82737c9b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 120, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 120, "uuid": "8436b1ad-0462-5d00-a84f-b89036d2dfc8", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5206519616767764, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8436b1ad-0462-5d00-a84f-b89036d2dfc8"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 129, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7a80eb686d7035011defb2775d32ad9566567e41b2d45c65e852af4299f7bf53.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/4c8da3f45f8baed076a2e521bb8fba92e91a8b5dd7edebcfc4a8f995d4f7ba23.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d2fa94370cd62981deb374c2e9d7dff2abdb59ac751a6da4a55b2514f87fd6de.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 120, "uuid": "9d1297e1-e7b8-5427-9dbd-e166c59fe58e", "validation": {"axes": {"B": 1, "N4": 129, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0028816862031817436, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9d1297e1-e7b8-5427-9dbd-e166c59fe58e"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6d60916011cde578ebcccad4448e657c00c49f53c0a4361a4532f739385f84a3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/75235ab00e75190806964cd429e4b81989d9a6862ce320e2f725d25a4fb11240.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/db56d04b702ccbf75b96ada555e5abb652f0d3b5250528cea10a76a0bc5b3995.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/6ed2fdca2277deb9b7778bb6d055d80bc2b65c68794701f3f45b9fab9ea5d976.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 121, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 121, "uuid": "4c2da6e9-5a34-5b15-9b7b-689d4389f09d", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 4.470348358154297e-08], "seconds": 0.002981953788548708, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4c2da6e9-5a34-5b15-9b7b-689d4389f09d"}, "variant": "decode_1"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/f5bd17366a02161d6946d489d3bce39189c8b5410978cd1e093801f90ac08e7b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 121, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 121, "uuid": "16d8d09a-daae-5853-834f-2875fb85a5ae", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8776261690072715, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "16d8d09a-daae-5853-834f-2875fb85a5ae"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/83b729789ab6435d41414a6e6ca79af5848a6ed331198457457d0a460f72b265.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 121, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 121, "uuid": "7713b508-dfc6-502f-be67-177b21b7c674", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.5179842039942741, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7713b508-dfc6-502f-be67-177b21b7c674"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 123, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/7e40c73d64119f7df96fd61dfbe3fbc7fe638280f7eeea5f49df9e6033757489.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c6c1e11fb47edd5b81dd10f3fa491a5a4df1adb5123ec582b830fd742b715144.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d2fa94370cd62981deb374c2e9d7dff2abdb59ac751a6da4a55b2514f87fd6de.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 121, "uuid": "61334221-7f5b-5c6f-8650-664972cff6d1", "validation": {"axes": {"B": 1, "N4": 123, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0028644781559705734, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "61334221-7f5b-5c6f-8650-664972cff6d1"}}
{"agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/09ccd29f0829f65025190c659895407160ecff9b3acd63ea211290aba52e59eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/13c28eed277a2c6df1df33376c37b8833286cf7318a269ae7e2ffbb56d61975f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"expert_ids": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/21c20f4cd20758f90ac00e0362996652d00ef5c67ad1f3fb19c4a96a72f52c94.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "expert_weights": {"path": "blob/workloads/moe_router_gemm_topk_deepseek41_d5120_e384_topk6/8b9dcb083ab7313fb9f1781b10bd9b9a891b0298f317af4dd05a3ef89963b9ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 0, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "MoEGate.forward followed by vision_topk", "output_adapter": "Apply the engine's deferred routing scale to its selected weights.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_order": "Serialize the unchanged selected expert set by increasing expert ID; permute the actual weights with their IDs.", "selection_step": 122, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 122, "uuid": "04b439a3-46d3-5070-9794-a00f75da3524", "validation": {"axes": {"M": 1}, "definition": "moe_router_gemm_topk_deepseek41_d5120_e384_topk6", "device": "cuda", "max_absolute_errors": [0, 5.960464477539063e-08], "seconds": 0.0030375015921890736, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "04b439a3-46d3-5070-9794-a00f75da3524"}, "variant": "decode_1"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n4608_k5120_g32/628150af8edcc16649d3f5c05d8b336d82cffd7a93d17e918f63b08595aabbab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 122, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 122, "uuid": "8ebe3e49-d085-5194-b8ef-3b26f685f3d0", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n4608_k5120_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.8652677489444613, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8ebe3e49-d085-5194-b8ef-3b26f685f3d0"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather all expert shards; discard rows marked -1 by the engine; preserve group-32 activation scales or expand each actual group-128 scale into four identical bytes.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/grouped_gemm_fp8_fp4_e384_n5120_k2304_g32/b3da1cf1cdb650bcfebd794569e2eaebef20f4f4fab374932924902f728d7803.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "SGLang DeepGEMM grouped GEMM", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "selection_step": 122, "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 122, "uuid": "94998979-421b-5dc0-a611-c2abe525b98d", "validation": {"axes": {"M": 6}, "definition": "grouped_gemm_fp8_fp4_e384_n5120_k2304_g32", "device": "cuda", "max_absolute_errors": [0.0], "seconds": 0.526547497138381, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "94998979-421b-5dc0-a611-c2abe525b98d"}, "variant": "decode_1_cached_False"}
{"adapter": "Gather real query heads from all tensor-parallel ranks; preserve all referenced physical pages and remap indices.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_4": [422126, 128, 1, 288], "allocated_cache_8": [234, 256, 1, 528], "axes": {"B": 1, "N4": 125, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "engine": "SGLang", "engine_adapter": "Pass actual dispatcher activation-group metadata to DeepGEMM; the pinned FP4-expert recipe otherwise hardcodes group 128 for group-32 activations. Tensor bytes and model architecture are unchanged.", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/680198fe3771d2425a4a5a9ef571afe84bd7651769606315499ea1cfa1028d5f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "deepseek-ai/DeepSeek-V4.1-Flash", "model_revision": "5254448103b6254fc646dacbca1cbbf20e9216c1", "operation": "flash_mla_with_kvcache mixed FP8/FP4", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "physical_pages_4": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/c84e6b21d4b3fc07c70dd0d693f32713087c292b3c08f67c71a412765402056d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "physical_pages_8": {"path": "blob/workloads/mixed_mla_decode_deepseek41_h64_d512_k8128_k4512/d2fa94370cd62981deb374c2e9d7dff2abdb59ac751a6da4a55b2514f87fd6de.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "deepseek41_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/deepseek41", "--host", "127.0.0.1", "--port", "30100", "--dtype", "bfloat16", "--trust-remote-code", "--tp-size", "4", "--ep-size", "4", "--moe-runner-backend", "deep_gemm", "--moe-a2a-backend", "none", "--disable-shared-experts-fusion", "--disable-cuda-graph", "--disable-overlap-schedule", "--cuda-graph-backend-prefill", "disabled", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.9", "--max-running-requests", "24", "--watchdog-timeout", "3600"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT": "shared", "SGLANG_DSV4_COMPRESSED_KV_LAYOUT": "auto", "SGLANG_DSV4_KV_LAYOUT": "v41", "SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 122, "uuid": "74f468ba-1533-57c6-af5d-746c40d970a9", "validation": {"axes": {"B": 1, "N4": 125, "N8": 1, "P4": 128, "P8": 256}, "definition": "mixed_mla_decode_deepseek41_h64_d512_k8128_k4512", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0029348651878535748, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "74f468ba-1533-57c6-af5d-746c40d970a9"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 554, "NP": 580, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4cf63375da0395c5e3ca60c230717e9871f5f29e75da12b0c26419ffc8f53224.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/73ad33758cdcfa87daec087ebfcfc107c69b20cbf67b700d9e25a530b24c0b90.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2c1d36eab8b431730d31aa1bef10ce7dc5fd354f16c63f2fcc4985f99d53d41e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/13f5956ada05905617cb6203847f0421597fb4a7f7991cdefc713450e82e205d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35009, 35393, 35457], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "5ec0101c-535c-580f-992e-53ed104e883c", "validation": {"axes": {"B": 3, "MP": 554, "NP": 580, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 2.7295097089372575, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5ec0101c-535c-580f-992e-53ed104e883c"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fa7a8de18c6b22bf96316fb9b0181e73b8da871a0ad38a3d4e5b515151dab9ee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1fee580281d8edb69d210d9cd5eda1ab1d6659fd1668040c0de3ea620cff6c55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35009, 35393, 35457], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "d8db059c-8442-54c9-a179-9c32a5bfd784", "validation": {"axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.3524620537646115, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d8db059c-8442-54c9-a179-9c32a5bfd784"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 554, "NP": 580, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7db563c6c14f1f2924b216dcab77b1b1d3cca39e111ed914d4cbc783d9c4f25e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0cab0ce8611a2f09ed38bd88313a408a388cb09cc964c756a91422f5b3bddf05.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2c1d36eab8b431730d31aa1bef10ce7dc5fd354f16c63f2fcc4985f99d53d41e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4a43c99ac52b583fb8fc1589d467b5326a4a536c1e2d73b41f0bddf054c0fb12.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35010, 35394, 35458], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "856569b5-29ea-57f2-bf3f-c04386f06782", "validation": {"axes": {"B": 3, "MP": 554, "NP": 580, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002611102070659399, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "856569b5-29ea-57f2-bf3f-c04386f06782"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/774d712e7e48f78ac5e5b83554ab746880e61736f433ba4562ba451ce97c0e69.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1fee580281d8edb69d210d9cd5eda1ab1d6659fd1668040c0de3ea620cff6c55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35010, 35394, 35458], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "6014fe0c-10a8-5a48-b19e-0d9f8c1a8ca0", "validation": {"axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.014078446198254824, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6014fe0c-10a8-5a48-b19e-0d9f8c1a8ca0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 554, "NP": 580, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0fe651720920a8435b665bbdeadc680d5948c1ab655fdf217dc1d2feac9c4d96.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6d3d06dd1d8c6d2815cc14f32feba043ab34b68fe6c6f64c96875a2e53f198f2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2c1d36eab8b431730d31aa1bef10ce7dc5fd354f16c63f2fcc4985f99d53d41e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e4c9713f199ca21c1aa7714287c7e3e7e5af45aaa5745b5a19624d446c0d6811.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35011, 35395, 35459], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "9de7158e-ac6a-5edd-b5f9-d76b4035e5aa", "validation": {"axes": {"B": 3, "MP": 554, "NP": 580, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0021531088277697563, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9de7158e-ac6a-5edd-b5f9-d76b4035e5aa"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b290f246938722443580cf4a0fc0265fe08aa79e3f4b2156fbae3a9c166b0b5d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1fee580281d8edb69d210d9cd5eda1ab1d6659fd1668040c0de3ea620cff6c55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35011, 35395, 35459], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "628abad1-1ad7-51f4-89a3-b46328ecfbe1", "validation": {"axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.013353561982512474, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "628abad1-1ad7-51f4-89a3-b46328ecfbe1"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 555, "NP": 583, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a78e13350c3be7202587e86cf6d6eeaafe5d2484498953fbdf6de35a3adac51e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f8edfe1afe6e3171984a6353e9690289953b102aec1458cdb6cc9f0f363f08b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1fee580281d8edb69d210d9cd5eda1ab1d6659fd1668040c0de3ea620cff6c55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ccb9aa42e25bbd7f011243eb611b0d09aeca4f4308d236d19313ef0a316ee234.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35012, 35396, 35460], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "801c787f-d6e1-52c1-a94e-98067513bdd5", "validation": {"axes": {"B": 3, "MP": 555, "NP": 583, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0021702852100133896, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "801c787f-d6e1-52c1-a94e-98067513bdd5"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6fab3e81677245661b065cc12fdccbbb7072e14d55f0992d28f03df085e4984f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1fee580281d8edb69d210d9cd5eda1ab1d6659fd1668040c0de3ea620cff6c55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35012, 35396, 35460], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "0651556b-26d2-50d5-ae02-a3d245081efa", "validation": {"axes": {"B": 3, "MP": 555, "NP": 583, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.013451595325022936, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0651556b-26d2-50d5-ae02-a3d245081efa"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8fbd9586832ca719a38074cd539c8a5caf824043a40e7c000e012fc1671375eb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/33be8e6194a7ba620c4eabdc53c920a4ad7ed533cfec62b99930665a1e92f2a8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/02cd3ad68cfd8910b254fc2bfe118a452824c7ee99c20de5e291f716482b7371.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35116, 35564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 107, "uuid": "90a96bc1-3128-515b-b504-89b863fccfe5", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0018524909391999245, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "90a96bc1-3128-515b-b504-89b863fccfe5"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/349234de24c16d0816939a58fc64b14aa0dd02b16d9440413348df532a71038d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35116, 35564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 107, "uuid": "37c2a312-b8ce-536b-8359-512c16d05bbb", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.012522632721811533, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "37c2a312-b8ce-536b-8359-512c16d05bbb"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/460c5fe3d161094501b0661888143eae365ae2417b4ef67b3443630efd0c671f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9663cc7893d1f7fa49ac67cc57d59c57f9874c2d6ef963c97060d1296a6bee02.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/40631212f111dbe7eb8bdce7ece13950744a069f7cca2d39ade04a979ae060f8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35117, 35565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 108, "uuid": "3182e335-dc46-58a5-9ed8-ac83e2ff5d7d", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0017699571326375008, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3182e335-dc46-58a5-9ed8-ac83e2ff5d7d"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d6b5c885ab26fe1091baf995ac3e6daec366347254c4a4c700c651b730ca1dc8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35117, 35565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 108, "uuid": "02355533-2a2b-53ea-8972-d21c4759ef00", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.01258636312559247, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "02355533-2a2b-53ea-8972-d21c4759ef00"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/620f5af150f8f61a5638e1cc28eeb4e367b68f34340bbf5db06fdcbf136c0678.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/33cefb890a5a8542731c217429c150c6967de2082dc079043437ba7e87a7fe06.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a371f7567d6176b74b0afaf8eb3564e6320c164f784ef111f86f8966a640d828.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35118, 35566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "a0eda304-9e60-5387-af1c-1d3019587cf3", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016879807226359844, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a0eda304-9e60-5387-af1c-1d3019587cf3"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/218da393a4c8657a492372e23e72f9ad948874480c92e933c94fe286e35cb018.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35118, 35566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 109, "uuid": "e72fed93-6056-5be1-adcf-218fe69a446d", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.012568847741931677, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e72fed93-6056-5be1-adcf-218fe69a446d"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/96a07280db95b2492d02388e6222bc08e24fb5536a9f048405042bad1bdf5536.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0da1ec03d5c7702f831b42976dfcc7bbcd62c40078dd71151135a95874626c0d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/39acfbeda768cfed8c54fafb87f69c0025cbf923a820d844c9cf2164719fc6a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35119, 35567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 110, "uuid": "a1483ba8-59e0-51b5-825f-498d28b6fa98", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001674720086157322, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a1483ba8-59e0-51b5-825f-498d28b6fa98"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8c157891067f463a132fe28c8016d5f52c926469c70cca070e35d66e68a6449d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c7d18c03303d79bbed0a69dfcb62c10222c3a4101d058d542297d1c139beceb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35119, 35567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 110, "uuid": "4fa2c1f4-fcf5-528b-837f-d77ecc86219a", "validation": {"axes": {"B": 2, "MP": 556, "NP": 568, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.012532952707260847, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4fa2c1f4-fcf5-528b-837f-d77ecc86219a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1bd04d086bfff1eb82c26bbbf60be81c663dda21c8baeff54ee6a243eee82baf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3853c70542327d32a784e44eb7523b9737897c996322bceb0d4f7f760b0bcdca.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/459f7a0eb9d4256b4ac9d1246f248ceddb9f057ce0438c6c12147e3b9e854e62.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 119, "uuid": "c1169039-d7e1-58d6-8371-f9a9d899bc43", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016057458706200123, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c1169039-d7e1-58d6-8371-f9a9d899bc43"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d64e6ed938f14efdfb227069ece7578f5bd7be0430696a70aff2fb805774e416.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 119, "uuid": "18f83437-e2f3-525e-9fb1-81bae92ff7de", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.015934498980641365, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "18f83437-e2f3-525e-9fb1-81bae92ff7de"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8a18318cfb88aa0f04d63d625628cffdc1c08fdd0682e721a1d0cf57caef5c89.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2a527f596c23b0b4af1435d9694222879177c1c9d05974a9e5231e3b3a423b35.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4e140dc03bc51a1133a0d4dde6c44199dfe1dc88572fbacf9fe532e2bec8f520.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 120, "uuid": "0c3a4592-10d6-578a-94fd-6262495a9e32", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0014676409773528576, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0c3a4592-10d6-578a-94fd-6262495a9e32"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5bebf8cdfce8c0b9cc459c0e3c4080fafdb9374a417b6bd9de795292f455054d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 120, "uuid": "894ca991-5930-57b9-bd19-230379823d1f", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.011793083045631647, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "894ca991-5930-57b9-bd19-230379823d1f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/40b999de5abe7bf058f76e1bae27d8259909ce9df2253b7e0845b56d306665ad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7bfe6ecc1f471efc43eeece30a026103c74ac98cace2026761e24cf2dc2d412f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c33bc97e10c60275a3342c4d8533e11e9308a3817e42d33888fcbdf81a2f358e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 121, "uuid": "c9e14aef-74f3-5159-8821-4518961dae11", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013808300718665123, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c9e14aef-74f3-5159-8821-4518961dae11"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/37d3deb7391339eeea51e3de3a5690e64ce3233e6eedbf4aff7ce89b336114a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 121, "uuid": "4675c248-ba29-5899-8cec-a33ddb29b53f", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.011808262206614017, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4675c248-ba29-5899-8cec-a33ddb29b53f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/84752e9ff3da58400d5478faa38088c7515325c9665e3036be05e0d2f8812cab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/29f71eeef366892ef13e7d7dc643d1fa444097001828e5a3edd85a5233ffc66e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/51e83982ee123dc76de5feb54104b509a5cd29c3a8e871f47082db3fc8070e69.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 122, "uuid": "d930f9e5-bc4e-51bc-aea1-d704d973ae6a", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013755839318037033, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d930f9e5-bc4e-51bc-aea1-d704d973ae6a"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3b95e616a1526bf3ed4cb7447f5e3aa68142fa5ef1e23b425bfcd64d9e9dd624.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_002_7bd2099b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35008, "locator": "requests[49].requests[0]", "model": "claude-haiku-4-5-20251001", "output_tokens": 120, "reusable_prefix_tokens": 34368, "reusable_prefix_tokens_256k": 34368, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86543.763, "uncached_tokens": 640, "uncached_tokens_256k": 640}, {"block_size": 64, "branch_id": "subagent_003_788ad06a", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35392, "locator": "requests[50].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 108, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86549.019, "uncached_tokens": 384, "uncached_tokens_256k": 384}, {"block_size": 64, "branch_id": "subagent_001_f552fe6f", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 35456, "locator": "requests[48].requests[1]", "model": "claude-haiku-4-5-20251001", "output_tokens": 126, "reusable_prefix_tokens": 35008, "reusable_prefix_tokens_256k": 35008, "role": "subagent", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207, "uncached_tokens": 448, "uncached_tokens_256k": 448}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/57b604f899c0c3ac230afe29fd1683877d9ddb9a3917840c9f1ba19b583af962.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 3, "label": "overlap_p50", "model": "claude-haiku-4-5-20251001", "session_id": "002001296e8a8c38ad9d7cc436d691afc602", "t": 86550.207}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p50", "scheduler_mode": "decode", "scheduler_sequence_lengths": [35579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 122, "uuid": "19a57667-9df2-5bfd-b1c5-d48ddd252b70", "validation": {"axes": {"B": 1, "MP": 556, "NP": 556, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.011921224184334278, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "19a57667-9df2-5bfd-b1c5-d48ddd252b70"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 505, "NP": 891, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8310fce1cca119e88208b48fcf997975975a392589f94456602a9cf91b5e3d30.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0ebd1c95e127c97060c8347c56506eb7dcf118e1f9a6dce22aceaa30773514cb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0a8e86b017faba0a5fbbbb923414d574b7f580dcf028e71f5b9006162c4d5ef8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f5a6f1eaf321e4f173ca5d6f5ae9ecb724d2678a9bae41639b46d0705189ee13.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32321, 2945, 2945, 2945, 2945, 2945, 2945, 6785, 6785], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "612c3615-466d-5b88-bb6f-c6e5eb6d3494", "validation": {"axes": {"B": 9, "MP": 505, "NP": 891, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003918652888387442, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "612c3615-466d-5b88-bb6f-c6e5eb6d3494"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e1f5c5fad6f8ca8bdf7600ed0e84399ba46c9a317fc06501c2f157aa48758421.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/241c060d836749170e261ce0662edb88fa8ec40ae395b2effe43dfe264893a1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32321, 2945, 2945, 2945, 2945, 2945, 2945, 6785, 6785], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "c748cb6e-0f90-55c1-95a6-36d5a01fb79b", "validation": {"axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.025503457989543676, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c748cb6e-0f90-55c1-95a6-36d5a01fb79b"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 505, "NP": 891, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/85cd4e64fcfbb5798551ec609c6f818bece342e828e4d653399b64578d0191ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d08fda4efa07843eae01ce1545582f93f60a0dc13c457a319bf08866b6dd0f89.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0a8e86b017faba0a5fbbbb923414d574b7f580dcf028e71f5b9006162c4d5ef8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/af79549d493c5558d4c81e8352a81edb0598c027689ce902048c8b3be34d53aa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32322, 2946, 2946, 2946, 2946, 2946, 2946, 6786, 6786], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "e49d38da-60a7-5cba-ba17-5b7fa05594d4", "validation": {"axes": {"B": 9, "MP": 505, "NP": 891, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0036089131608605385, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e49d38da-60a7-5cba-ba17-5b7fa05594d4"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a4a5e04161fa66c9af568ec3579b3a11004d87c5600271d5f5497074856a0f70.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/241c060d836749170e261ce0662edb88fa8ec40ae395b2effe43dfe264893a1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32322, 2946, 2946, 2946, 2946, 2946, 2946, 6786, 6786], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "4e953b18-bcd0-540d-b1d7-45dd1b5e6350", "validation": {"axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02208658494055271, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4e953b18-bcd0-540d-b1d7-45dd1b5e6350"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 505, "NP": 891, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8ba66c0c0eddfe17681b4a3b86b43785470cd505d1ec17789e47e13707555bd2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e410a7ab4b90dca55c96e89005755b4869652ff2299fa5b10eba66990024cc6f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0a8e86b017faba0a5fbbbb923414d574b7f580dcf028e71f5b9006162c4d5ef8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fd9a8eb6a34bf9fb56094f0034877f8bbf0aed7cdcabfb2ac34f7a8582c4a556.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32323, 2947, 2947, 2947, 2947, 2947, 2947, 6787, 6787], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "218d219d-86de-5358-9b76-36def5a157f5", "validation": {"axes": {"B": 9, "MP": 505, "NP": 891, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0034547587856650352, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "218d219d-86de-5358-9b76-36def5a157f5"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3e682fdb86e14587d4db4a09938b8836998c6117951d9fce6cf16ceecd11444b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/241c060d836749170e261ce0662edb88fa8ec40ae395b2effe43dfe264893a1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32323, 2947, 2947, 2947, 2947, 2947, 2947, 6787, 6787], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "c6f99f1a-63a5-5f7f-86e2-8693a984e137", "validation": {"axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.022344077937304974, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6f99f1a-63a5-5f7f-86e2-8693a984e137"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 506, "NP": 900, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7e86418b7892ebfab6452b47fa12e2d6c87420445fdea8bd2e5733278c4d3f67.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c1f5c3a4e552a50ea47414c9d601e65cead5b3367a4beb9cd974b007a99c24af.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/241c060d836749170e261ce0662edb88fa8ec40ae395b2effe43dfe264893a1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8de2416f5f1ae33566f7ef8471e4f2194577fa4a3d441654de304f8949141e55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32324, 2948, 2948, 2948, 2948, 2948, 2948, 6788, 6788], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "92278181-b1fd-55d0-bd61-7e2ba5e5a3e9", "validation": {"axes": {"B": 9, "MP": 506, "NP": 900, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0036361338570713997, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "92278181-b1fd-55d0-bd61-7e2ba5e5a3e9"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fadc5776671a41ef07fe3cdfb09aa59c188a2b217ae876bf4e4d582df93bfdd7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/241c060d836749170e261ce0662edb88fa8ec40ae395b2effe43dfe264893a1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32324, 2948, 2948, 2948, 2948, 2948, 2948, 6788, 6788], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "4181544d-7887-5557-b193-9acac92a66fd", "validation": {"axes": {"B": 9, "MP": 506, "NP": 900, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02207024209201336, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4181544d-7887-5557-b193-9acac92a66fd"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7189e842630abaafbfecaeedeeac994c75cb6d71f37ef1453690255e6e9dc8e9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d86dfc2a90618984aefc2675143b077062ca4a8388333ae73297cf0d6805c3c2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fc24738b23014be9679d25a5f73fdc86b48569d2c10d6aebcab35cc981bfb696.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32681, 3305, 3305, 3305, 3305, 3305, 3305, 7145], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 360, "uuid": "2eff4b84-9fd1-56b5-b0b3-e1c266e5b9a2", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003435838036239147, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2eff4b84-9fd1-56b5-b0b3-e1c266e5b9a2"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f32e4628dae73b37524a0a31a9ca2f1a0225972b7dd172f51ef70c725e82ecf6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32681, 3305, 3305, 3305, 3305, 3305, 3305, 7145], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 360, "uuid": "685cb0b3-495f-5b84-afb9-5b0ad540901e", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.022736758925020695, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "685cb0b3-495f-5b84-afb9-5b0ad540901e"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c856e213cddf57061083f1ebfcc18dbb4f74878ed1dfd5633f94e8dce44aca18.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fcbaa1d3a2fa963a90af267eccab783f76c009ac1059f0e21c84e4c2bd0aa6f0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d4f9e07b5348e09138c4eac3710d77f03068bac1290e6baa03b7c60dd14f0091.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32682, 3306, 3306, 3306, 3306, 3306, 3306, 7146], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 361, "uuid": "36e47cae-8617-531a-94eb-2905d6e8af00", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003370008897036314, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "36e47cae-8617-531a-94eb-2905d6e8af00"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3a43293029675b4dde538cfd34ae663c4801f50097bbe272199a72fded3a7ab6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32682, 3306, 3306, 3306, 3306, 3306, 3306, 7146], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 361, "uuid": "d266616e-9234-5297-8b29-c9df01600a55", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.022339119110256433, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d266616e-9234-5297-8b29-c9df01600a55"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cfcc84bb7e12e37478ba375efa5113c77d41cc51d192689eccbbf75b21b5cfe8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3a2ddab5c8fc80023a121a804e2a2e04960df8fe4e35e5598291ed66fa1f9849.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/da596dadf83a3cb322a3398438d89c5441e6655f8887a4dee51bd67cde9b43e7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32683, 3307, 3307, 3307, 3307, 3307, 3307, 7147], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 362, "uuid": "c99d8cdf-15ac-5f6d-be07-c295458db43e", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0031792819499969482, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c99d8cdf-15ac-5f6d-be07-c295458db43e"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/37d1ce0cc0d49c83ab820b66ae617792e2a005e2861bcecb9e4f186b46e27caa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32683, 3307, 3307, 3307, 3307, 3307, 3307, 7147], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 362, "uuid": "a5fd3fec-d855-5843-8ace-b6e3d459c626", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.022309234831482172, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a5fd3fec-d855-5843-8ace-b6e3d459c626"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/36804b9e94e50297146444643a218f7ffceaf2b185179650d7c932b0acabedec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cc4887f4e08185111b3220cb351f37a1d439fe4f825f6c11d78400a85818c3f2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f6a7a734e56f35eb72eacdb0371003c704540b74b9343e2777cab158ecb95376.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32684, 3308, 3308, 3308, 3308, 3308, 3308, 7148], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 363, "uuid": "9b61689c-d8a3-57b7-b0cc-a3186353895f", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0033697891049087048, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9b61689c-d8a3-57b7-b0cc-a3186353895f"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fc39d8445112aaf17fbffc9378aba02f0ec21bfe82f9b70214b9979ab0895626.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/06dd72858a1703457914440d077301459b134a5024dc0088771bd8c1b1699e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32684, 3308, 3308, 3308, 3308, 3308, 3308, 7148], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 363, "uuid": "f474826b-c0b5-5266-8811-450bd32ceb86", "validation": {"axes": {"B": 8, "MP": 511, "NP": 935, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.02231942117214203, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f474826b-c0b5-5266-8811-450bd32ceb86"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/48beaadc3c5a8dc03f096a930fc2f571d566346e3908656936042cfc4883135e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/be5f6ca2fdd57129ca62610a9ad1d9db0f6ef1d4267eebd74dd1f1e0908d64cc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f05d16a42961cbf755d5c1f9b19ce563be136a0b9865c081bc15aec494255e48.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32783, 3407, 3407, 3407, 3407, 3407, 3407], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 462, "uuid": "28353c79-b9c0-5ca3-a37c-646b64ed5917", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0031166430562734604, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "28353c79-b9c0-5ca3-a37c-646b64ed5917"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/23a84790b40d95719433c7bc0aede6b55332b580abe301706af4717b00ed0568.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32783, 3407, 3407, 3407, 3407, 3407, 3407], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 462, "uuid": "41b6e76c-f7db-5aea-bf47-854a33c9b15f", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.019836103077977896, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "41b6e76c-f7db-5aea-bf47-854a33c9b15f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0e57ed58caf5c2cfb58e043bd077421b708465ba7e4557392263b37955bf8bd2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2fb45d395b2fef6be92cc7390a4594ddad96b626ec55042cdcf68a656448075c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/770d4639029fe50fff0d5ccde324001b96aa53f34df0802047f010042ed7488f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32784, 3408, 3408, 3408, 3408, 3408, 3408], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 463, "uuid": "f5a5b78e-10aa-5910-92fa-6bc66cee8bb0", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003096809610724449, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f5a5b78e-10aa-5910-92fa-6bc66cee8bb0"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a425c8cfdafa153abc0fa09fd3541761826d1ceb9b443b527d70894ff5566c5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32784, 3408, 3408, 3408, 3408, 3408, 3408], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 463, "uuid": "70d1ffc7-e0e1-5aa5-97c2-976968159988", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.020015966147184372, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "70d1ffc7-e0e1-5aa5-97c2-976968159988"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/97dab43f8de37e7b28612531bdf5ef08c36f98eb4769697e926bb3245c8c7f83.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/12539197a53c37f9f9369e33645f9ec560bf37f5641f770ffda70d22ec840bd2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/abee2ca71a2db0dc7045814518326d8a4ec0020eee68b0e0e67f1936318d57a6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32785, 3409, 3409, 3409, 3409, 3409, 3409], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 464, "uuid": "1be50d68-312a-5a5d-8381-901fecda6693", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003051979932934046, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1be50d68-312a-5a5d-8381-901fecda6693"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e77cc166751b0c5fed8742627fba2e9f2e85a84b49acd8feabb76cf8144a9d6c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32785, 3409, 3409, 3409, 3409, 3409, 3409], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 464, "uuid": "137a646b-3fe3-5037-9995-fd58a4cb1486", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.019908771850168705, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "137a646b-3fe3-5037-9995-fd58a4cb1486"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/889aaa7ab5e3dc4c8aed54f4a83119a5b3ac1baab767a40d3cbaa2d9a7cfa064.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8294f255f419f55f8f4ec4833712a84c0d2ac2cd50d290156f6a333614811b77.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0d1155b0ac712787a338962494825deacc59b75ccf36d37a8275852ca11a8543.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32786, 3410, 3410, 3410, 3410, 3410, 3410], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 465, "uuid": "84bbdd35-cf32-5eef-b35f-647ae858aa4c", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0029254453256726265, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "84bbdd35-cf32-5eef-b35f-647ae858aa4c"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/67d6015512e27a63357b0cd66782d1b3b716e455eff0f9ae8e3802ca33d2c8fb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1d28a3355541641bf6199017097b6c15eb3d72615c1cab8705242412c4a60115.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [32786, 3410, 3410, 3410, 3410, 3410, 3410], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 465, "uuid": "8d24707c-f5e2-5078-9acb-620102255fa0", "validation": {"axes": {"B": 7, "MP": 513, "NP": 837, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.019903571344912052, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8d24707c-f5e2-5078-9acb-620102255fa0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/319cda21a7012d2374230555058807d8bfb72f260a76d04dce6d29b2bdca03c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/86828f6d41094945e0f8f9acdc69ef3f33e2376c66403518976e44d44083dba2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4614d251cf2a8b381ddd27e9f00662d0c91c893d374a98b8e0a3bb7073d3d185.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3576, 3576, 3576, 3576, 3576, 3576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 631, "uuid": "453dca61-6af4-5497-bbbe-7449d3993f83", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002551883924752474, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "453dca61-6af4-5497-bbbe-7449d3993f83"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8bb496efb0dccd0db167ba279479291e62fbdd624e31f941c11643efb6161b79.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3576, 3576, 3576, 3576, 3576, 3576], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 631, "uuid": "5013fbf7-df16-505d-912c-d20c3a8260b1", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.014230554923415184, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5013fbf7-df16-505d-912c-d20c3a8260b1"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/318fcc1331f9b3486d14031ac5e8962c46181296f53b3fb35c771667d754013e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/22d2ffcd26161fe0bea1547f3493c904747743edc413cc6d5a021b5773950a1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/56e97f1b6a5b5b52a6cbfccd7b93bae945fc88635d9af77c891ac24a78360446.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3577, 3577, 3577, 3577, 3577, 3577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 632, "uuid": "1eef615a-2a43-5200-a66d-93cb5fe0e925", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0024990346282720566, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1eef615a-2a43-5200-a66d-93cb5fe0e925"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5cc166ca40586a87175887f2344ef7133db7a1d7cf7814c3e589ce7ee06d7e11.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3577, 3577, 3577, 3577, 3577, 3577], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 632, "uuid": "e488e086-e613-5966-964b-bfba1b917888", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010613103862851858, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e488e086-e613-5966-964b-bfba1b917888"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/dbec96c08084618336ee10e5d2b6cad5ef6734859fe31fb5c445bbc7745313d2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/627571d26c1e58b776d9e5bfc7f8a723cda1e16331fcc732427df3ead33bdc24.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9ec7d980d0b63ed4456a60a27ea55fee25ea89fe15694a8a48c0505e74749f0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3578, 3578, 3578, 3578, 3578, 3578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 633, "uuid": "dc98e089-3527-58dc-9948-80fb81ebd15d", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002386264968663454, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dc98e089-3527-58dc-9948-80fb81ebd15d"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ea854328c1b609f02f90fedd99699f29ab70ae6eb93cfffd7bc035ae80b77544.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3578, 3578, 3578, 3578, 3578, 3578], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 633, "uuid": "47f080c6-633d-5985-bdba-00a92cf3e4d3", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010529843159019947, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "47f080c6-633d-5985-bdba-00a92cf3e4d3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f56fe3e761036434d3c4a2be54e286e9a7da0a2c20465dacb5c01d16b65cbd92.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e5bd0c67fa92ba2103cd735f5a27bced946c6f3b75cd0d7f24cd4c55bb47223d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/430ce8f1af3e7461f39a48fa27c7dcc730bc8931abeb0f24d4b4020cb961b03a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3579, 3579, 3579, 3579, 3579, 3579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 634, "uuid": "7851c260-feab-5c47-8ec6-de5a79587efb", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0024084951728582382, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7851c260-feab-5c47-8ec6-de5a79587efb"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d948ad22ea5ab6501e4f9fd504ed56aef29ae048882811a1fc81f0800e2b8535.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4c863b3b6681585efb91726722a0362b2058ea10fc0ddbbc1e422068ed5627a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3579, 3579, 3579, 3579, 3579, 3579], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 634, "uuid": "27ba589e-c627-5924-ac17-274a65a3b8ab", "validation": {"axes": {"B": 6, "MP": 56, "NP": 336, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010504536796361208, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "27ba589e-c627-5924-ac17-274a65a3b8ab"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/81a0b721b2170a075060fbbc346e1d022e247a3f37573f8133889d708ff9be3f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a8c00c12bf6b97af5bfe7d1f0f41e6228ba622d99788fbac8347c9306e67c38d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/46399eb7cc2e3f8070d1f7203297e3a05b28357546935303066a80c3a9e5281a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4024, 4024, 4024, 4024, 4024], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1079, "uuid": "885b2098-4ca8-5a7d-8eb8-7277576a11f7", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002264623995870352, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "885b2098-4ca8-5a7d-8eb8-7277576a11f7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5268b0662a9298bb9398c06e8240736a12bc5ec8bb2941a4f06247d6130d00ed.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4024, 4024, 4024, 4024, 4024], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1079, "uuid": "ec59c720-3793-5e97-9532-fb637271838f", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.00960067193955183, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ec59c720-3793-5e97-9532-fb637271838f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/386df3d3d9b15d38087e47d9599267367674caf256b35b3a361eaa98bcca2418.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ba20444bf9e5433a60b39d08a4b04f77a13420c53eb5e76968597c0ea51d1153.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a875d3a8677d47dd1493986d2332a103d19db7c3218a707d428a3c12bfffaad1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4025, 4025, 4025, 4025, 4025], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1080, "uuid": "fd129989-97c3-5a5d-a922-c725b21aeafa", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002215445041656494, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fd129989-97c3-5a5d-a922-c725b21aeafa"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b00c5e0192fd51c29bc8b9b4e17cc222eef93472d07133128ccf244380f095dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4025, 4025, 4025, 4025, 4025], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1080, "uuid": "760b5dbe-900b-5ad2-80b6-5894c43b7316", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0097688059322536, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "760b5dbe-900b-5ad2-80b6-5894c43b7316"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1568c02113fef421374cc20a4e864e9bbb0d93d3e99d6d574b5797baeebd76d7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/edb7679a5c2d87fd1eeb9191b615c9b26e0db5845115965c7239b41eb2d3fa8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/33709e15c863fa1d04f8ca9580a0bfcc02a94ff45f93a67fbbcfdafddf2420a0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4026, 4026, 4026, 4026, 4026], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1081, "uuid": "9fdceab3-5a88-52b9-97ac-7bdffcf254b6", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0021811770275235176, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9fdceab3-5a88-52b9-97ac-7bdffcf254b6"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/eff3d1f7edbb8ca26a2687895471996051b7d688c3758fb23aaa5d0c32f65408.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4026, 4026, 4026, 4026, 4026], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1081, "uuid": "a054cc78-40eb-5360-9a14-e9b6ecc6fbcd", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.00964914495125413, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a054cc78-40eb-5360-9a14-e9b6ecc6fbcd"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/36463d04c170f6d55ace835d2dc8f85bb993f7506a3bc2bd441acf6d4e159bd9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f94210bb6a7419678fed9ec71f701df0fe6d6a30431df0df52d325c36cc7a079.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/876cc738879ac2d7e2ea5d5db59a65d7b84ea294e02c69ff3a349f72b513f8dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4027, 4027, 4027, 4027, 4027], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1082, "uuid": "cc0aa9dc-980f-56cc-88d3-cbaab30edae7", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002171626780182123, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cc0aa9dc-980f-56cc-88d3-cbaab30edae7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a2970a6f9c88bd3135989ae98da812ca2c97ea4fa1704a129f1d1fffb0e56a5a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a75cecb7cd4e90851dc8b28a52d54acd8032f5cece97fd17a1a44c163f04eff6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4027, 4027, 4027, 4027, 4027], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1082, "uuid": "86a30c14-5b7e-5111-88e3-8a1d8a1e32b6", "validation": {"axes": {"B": 5, "MP": 63, "NP": 315, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.009706938173621893, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "86a30c14-5b7e-5111-88e3-8a1d8a1e32b6"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f1290a83db0474216f7dfbb3241a1102ca1ded15d74ab884a50007a03c0e8e7b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/64c1e20bdc0c101e2f1d812198bab0efe9f7ca8639e1771ed6d26d65c50a19a7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8e82d322195608d71e96c3d27b389923c66fbb0e59ac495b5e57fa1d3b970045.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4092, 4092, 4092, 4092], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1147, "uuid": "b1ca8af2-4a8b-5624-9b42-e78941d4961d", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0020687589421868324, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b1ca8af2-4a8b-5624-9b42-e78941d4961d"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9d4058132f0781726f2d7eec74cf42e4cc72c1abca18402c9d21022e236f79dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4092, 4092, 4092, 4092], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1147, "uuid": "789332ef-12e6-5999-9405-3fecaee95851", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.010785181075334549, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "789332ef-12e6-5999-9405-3fecaee95851"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ff46de81efeb4fed84e57ff064a07252dad3a5e69438600fee95d937f315b3df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/17477b7913a5ae20d4a0c8b40bac56bfc4d204bfa1f6fa4c0294d7c20b37f9fe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/faa1ae6321ff1241f0fa57ebac42f7a27695989a67111488ab7ede0ccb33be16.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4093, 4093, 4093, 4093], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1148, "uuid": "91d84d5e-f04e-5eb1-a072-1c4fac569480", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0019682971760630608, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "91d84d5e-f04e-5eb1-a072-1c4fac569480"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8eeb363fda7c301335631142f1fd4790eaf1763d6637b72f0f82d9878dbbecea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4093, 4093, 4093, 4093], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1148, "uuid": "33599fa7-eafb-5399-b847-8d714567d7b5", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.0081334481947124, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "33599fa7-eafb-5399-b847-8d714567d7b5"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/091f766f76a7a12eb709eb0656e1306f020faffde0d42703ed219f6a8b7bea55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ba39a0d3fa2e8d5c44f4957e88cbcfb4fed25bf9d6c27d22def8d9b46d5c7ded.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/63fc815d9250df78d463dbc11ffedb8156788d9afd2ec31b9f0999827ca49ca9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4094, 4094, 4094, 4094], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1149, "uuid": "94444ed3-9566-5321-877e-f2b6297fcc32", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001919562928378582, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "94444ed3-9566-5321-877e-f2b6297fcc32"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/af73a925c528839b9c1a5af057a379df9e26a82128e44036017639119010cb01.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4094, 4094, 4094, 4094], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1149, "uuid": "0eae3d3c-25cf-5428-9fab-5d1210c913f2", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.008021803107112646, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0eae3d3c-25cf-5428-9fab-5d1210c913f2"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/74b0505f46133dfd7b5127608efa8b2474f9104141d1b9d95d70bfc303a97e07.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/28854bb7d59b8871802116db3819643445d4d62b6dc132686a85995354b19f43.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d859d93a6dc0ea74cfa6096e27853f072aa20f393c0009e57e63923606f1f961.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4095, 4095, 4095, 4095], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1150, "uuid": "1af52f53-ce8a-5ca8-8700-2d4ce4549f86", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001903959084302187, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1af52f53-ce8a-5ca8-8700-2d4ce4549f86"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ece05a16b80853b210887fbda38e7b47f7c0084fbc2ca1a61c0ff158ca433b3d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c803878483cf29720962c18b7765c8c3f3e29ce43ce6b69a4199ea90ea03248d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4095, 4095, 4095, 4095], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1150, "uuid": "de503f7a-7745-5927-80b0-e1ad27da211b", "validation": {"axes": {"B": 4, "MP": 64, "NP": 256, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.00801469711586833, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "de503f7a-7745-5927-80b0-e1ad27da211b"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 66, "NP": 198, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/024952216d9b3f7e9b77f3e072814fed9ed821bd3fdebdbb73ddbd26af30e181.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f704405588538c8dd964e99c23e3fb8cff928d14d3290c6ca04581071fd2068f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/62521783ec51f694a694be3e4ece32f0bb0e09d21601169f558f9f60ef287b8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2debe839d08280d0dc55d40f2c1b1a28b0ffe981c367a54f4ca5f63ef5849658.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4193, 4193, 4193], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1248, "uuid": "12dfdb1c-344f-514c-82bc-f93ce43110c7", "validation": {"axes": {"B": 3, "MP": 66, "NP": 198, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016703121364116669, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "12dfdb1c-344f-514c-82bc-f93ce43110c7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 66, "NP": 198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/83f5525b6530a054d8bed032ec47bbf39932507933689f1a90ede00cf8b50fd5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/62521783ec51f694a694be3e4ece32f0bb0e09d21601169f558f9f60ef287b8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4193, 4193, 4193], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1248, "uuid": "96040f16-5f7b-5852-812d-2da3a0c97e8e", "validation": {"axes": {"B": 3, "MP": 66, "NP": 198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.006414942909032106, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "96040f16-5f7b-5852-812d-2da3a0c97e8e"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 66, "NP": 198, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/472b21e0b9860624642abb089a2eab9386205062f1ef8a58d712af2187cb07fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/eb5880e4fa92976e071699b4e413b23401847eb5b9db5ed701e5e770726982cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/62521783ec51f694a694be3e4ece32f0bb0e09d21601169f558f9f60ef287b8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bc22a86df06287a9fae2e32775114903ff6efd933a051ff2f9194d2d1522ffe1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4194, 4194, 4194], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1249, "uuid": "d5431d8b-f359-5817-9088-f0614b68ca5e", "validation": {"axes": {"B": 3, "MP": 66, "NP": 198, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0017431299202144146, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d5431d8b-f359-5817-9088-f0614b68ca5e"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 66, "NP": 198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3cb370f264378d8543aaaeb5d08138be5715770e54422a4c0a1f524233ec495f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/62521783ec51f694a694be3e4ece32f0bb0e09d21601169f558f9f60ef287b8e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4194, 4194, 4194], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1249, "uuid": "9368a3c6-bf54-57b6-8f0a-1cc85c55945a", "validation": {"axes": {"B": 3, "MP": 66, "NP": 198, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.006466732360422611, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9368a3c6-bf54-57b6-8f0a-1cc85c55945a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6cbebd4585b0f349ed20afa8ed2d79278357496e9aff5ae59a3b33b9327d2d10.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a842138fadca2b6c0a6180b6868bf08f8644fe479c9bd82ad27225ebf174c20c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/46554a2ab23daa912ff46b8f0991399d122d6d461f338101d51b94a726de6b85.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4195, 4195], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1250, "uuid": "b36bd510-ddad-552a-850a-ac8ebb157d85", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013859779573976994, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b36bd510-ddad-552a-850a-ac8ebb157d85"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/77c6684cc55dd20ca4976953adc3e334179b19bd1227638805c1d3ee1a80446b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4195, 4195], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1250, "uuid": "ec1c6f87-c372-52c1-a1ad-7871fdeb9061", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004715308081358671, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ec1c6f87-c372-52c1-a1ad-7871fdeb9061"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/07ae9c41f0f37925c43e658a13b881ba22d59e282ebfbb1bbd1ce93bec6bca2f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/222e7c45b91f891f10f76116c746ad22e5aa3a5992d2295c560d0a780175630b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a948da4785920b1d1d6cc509f4f1c2acfbd7a4fa57fea690748acb260e666c40.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4196, 4196], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1251, "uuid": "364079ec-117a-5bfc-9dba-320012d7f1d7", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013824538327753544, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "364079ec-117a-5bfc-9dba-320012d7f1d7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/dc1a4d608d2a2d530094f00a8f0b1a97160f992c4a9fad529443639a87046339.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4196, 4196], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1251, "uuid": "676404fc-286a-51d4-bf6e-a8c3352340e5", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004721508361399174, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "676404fc-286a-51d4-bf6e-a8c3352340e5"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/00784e8012f131bd31a63d111844597d7452844348a70c84541e8307987de965.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/de11c90423079f21c71de1ddb78aea228352e265ee67d2d94c47d3645eb17ab2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fc526dfb658f1c517a4b6e5817a4d97423e792359ff12b382f83fc2be6ae16c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4197, 4197], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1252, "uuid": "3f78c02b-e626-5976-880f-b2eb0e37c257", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0015103379264473915, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3f78c02b-e626-5976-880f-b2eb0e37c257"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2ed0572a2753ee47424e6e6929cfc968b01c0cc75ef4da324a0ec3181f270068.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4197, 4197], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1252, "uuid": "4e9f56ae-bdcc-5273-8c3d-590ba7b00473", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004723663907498121, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4e9f56ae-bdcc-5273-8c3d-590ba7b00473"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/23740a4b9d836bdc4b5d72af4d1cd239ac9fa35316692d489819d66a06790130.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/19fcb2712f44017311549bd35fd7f7378c3d8d46e98aaf8fb6b21f0277c80476.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/56b9eb3a13a10e53adac72c51b0f6dfae0a10ae7ed99ae98f9d7dab3f16cff1c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4198, 4198], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1253, "uuid": "e0556925-e7eb-554b-a898-5204eb8303c3", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013251728378236294, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e0556925-e7eb-554b-a898-5204eb8303c3"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/58b3871a75b265b468397f9b7e609e943dfca932427144184b9c0c9360a2ee3a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e3be91d558c4b1cb56b0945ed245e66e8747ebf66735553c403e50044f7b099.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4198, 4198], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1253, "uuid": "91c2c92b-6ee4-5f2a-a761-e4f2b5bf6856", "validation": {"axes": {"B": 2, "MP": 66, "NP": 132, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004705972969532013, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "91c2c92b-6ee4-5f2a-a761-e4f2b5bf6856"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/839e822766db4f16767db056079530fe1f8f7c0606dcd04218f14a168a2d8f80.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7c3ae53226f719bdc13eb8ba5bd25ea72072f37c4cfd3d4954b660e573a82a4b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/65aa8e90b95eacd0ac8977ce194e1a140e82008fa0f8b0a685b0582306cad3ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1619, "uuid": "d5328518-1298-535d-81f1-d7632f2d5877", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0010986770503222942, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d5328518-1298-535d-81f1-d7632f2d5877"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ffee1be04e8b97279cbbbf1053acf0f10a0a2c8fdab57a6c3d47366b8245380f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4564], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1619, "uuid": "36162f12-70db-5202-85cf-903d6cc1eaf8", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003484856802970171, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "36162f12-70db-5202-85cf-903d6cc1eaf8"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e1b71effb9721bd1149fb4cddf74f91b35926bc974daa12aeecca4efc8daeb79.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ff0811653e3658adf26bda50ca70d4cd0f5cf364eaed0de8c27ee3f6a69d24c7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/46c63b5663d5c757419e612d94360d077b96d60499489c217e160a4bd38070cd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1620, "uuid": "f399b922-04ca-5b87-81b6-93e946021dd4", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0010723560117185116, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f399b922-04ca-5b87-81b6-93e946021dd4"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/49d1d145c7189112c415c98515b40389b408d9a54c15e619a081495817cc8be4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4565], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1620, "uuid": "f3167f71-4335-56f6-a51a-bf66932c243b", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003137522842735052, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f3167f71-4335-56f6-a51a-bf66932c243b"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0a8fde245a563c7f6fe4b272c86c9d112de30188e77a5f1712723a73075e14d9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/15017b922346a72364cb976fb6a350a573f647c9be956fb27e1d1a0bd23703ca.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/93f86417deb989e192824a9e4c577ff7221b3b3ba2a2ca28a0fa27157ab859bf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1621, "uuid": "9f19502a-3868-5b69-b099-c21210d7a27d", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0010344432666897774, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9f19502a-3868-5b69-b099-c21210d7a27d"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d1a8333f5ee5e3703b978ad4880f6348e00b3504e9d110613e227b17f94458a5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4566], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1621, "uuid": "d1e9ee09-0fd3-56e8-bd59-8cb0f6175db3", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003161177970468998, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d1e9ee09-0fd3-56e8-bd59-8cb0f6175db3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3da2b7348750730ec82845fa844c4a235bab76c9484bffea86768e803032654f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e4d713564562ce6b0e480f06eace42a26e1049859d1f9d30526c81f307cbea85.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c40c229ae8c7643c3faae82f620b28f09bb279b7c102586865e33315ba6082dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1622, "uuid": "0a66a9bd-1b59-5940-bb1a-4aa25162d044", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0010473579168319702, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0a66a9bd-1b59-5940-bb1a-4aa25162d044"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f5107fa038ff6c99dd63a41fd4c542edc237eb0e1be57d50fb8458d642b565bc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_011_37a5e0ee", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 32320, "locator": "requests[171].requests[3]", "model": "claude-opus-4-8", "output_tokens": 632, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17266.627, "uncached_tokens": 32320, "uncached_tokens_256k": 32320}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1080, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17268.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1810, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.611, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[172].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1148, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.67, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[174].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1620, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17269.866, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1249, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.499, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_013_a9e0df34", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[173].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1251, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.518, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_012_be174c2e", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[172].requests[4]", "model": "claude-opus-4-8", "output_tokens": 361, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17271.542, "uncached_tokens": 256, "uncached_tokens_256k": 256}, {"block_size": 64, "branch_id": "subagent_014_30f64472", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 6784, "locator": "requests[174].requests[4]", "model": "claude-opus-4-8", "output_tokens": 463, "reusable_prefix_tokens": 6528, "reusable_prefix_tokens_256k": 6528, "role": "subagent", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156, "uncached_tokens": 256, "uncached_tokens_256k": 256}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/78198dc1abfc21fd208338892ceb203dc446fea91a41f0a4fe1e800c9d54d63b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 9, "label": "overlap_p95", "model": "claude-opus-4-8", "session_id": "208d6e28928b0b3d3b33db3c690173d15be3", "t": 17273.156}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p95", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4567], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1622, "uuid": "803d0722-83e4-5ef6-b066-afc6373d02b6", "validation": {"axes": {"B": 1, "MP": 72, "NP": 72, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003141102846711874, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "803d0722-83e4-5ef6-b066-afc6373d02b6"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 23, "MP": 516, "NP": 1527, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/03c4198ba17e8be51bc5ebecdd34df873f25f79e228e27b8bd6d3d476e25feb5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/dc8aca90a5506b02799fc8ea300fb814c12ecaf470b39f32b9fdb90df5a6300a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0f7ee7a4a5fb4a6bcd05f8b498a277af3123eb2acdf0bd715bb6d6262bbced63.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a557602f43714667763d7ffd72fd7478e05335beeb53abf4187b4b0fd70d6cfe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 33025, 2945, 2945, 2945, 2945, 2945, 2945, 2945], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "fafad0f7-11c0-56fc-bab2-8313ee84f349", "validation": {"axes": {"B": 23, "MP": 516, "NP": 1527, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.007311367895454168, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fafad0f7-11c0-56fc-bab2-8313ee84f349"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fe5fdd5f9dd3d1109d9280ddd2208db18ff56ff53690b1382cfa2a3c48001384.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b1b5b7846f1dab5cd1d091230424d61edff6c8065ac23190dfe9c55238fcd92f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 2945, 33025, 2945, 2945, 2945, 2945, 2945, 2945, 2945], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 0, "uuid": "f89635f9-8d97-5102-a44e-8a249e9410e0", "validation": {"axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.05967831099405885, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f89635f9-8d97-5102-a44e-8a249e9410e0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 23, "MP": 516, "NP": 1527, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bc3c48c8e2df2a5200db8dcde45ad39ddf57b784fae0f8b6fcb35f0370b0b29f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8619de6221c2ea23d807d49f88dfb10ea3f5f6609c59c966db732d4d9744544a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0f7ee7a4a5fb4a6bcd05f8b498a277af3123eb2acdf0bd715bb6d6262bbced63.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e079cda1cb7293cb004dc6eac5ab11dbd95730aeee91c962e026e131b1c4f216.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 33026, 2946, 2946, 2946, 2946, 2946, 2946, 2946], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "80a299de-14e0-5632-a70e-b68431f49d5d", "validation": {"axes": {"B": 23, "MP": 516, "NP": 1527, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.00724219111725688, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "80a299de-14e0-5632-a70e-b68431f49d5d"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3f3fe9922ee92763fe3f2496a2b98919f32e4aa7aec7202a66c7dc77b0355657.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b1b5b7846f1dab5cd1d091230424d61edff6c8065ac23190dfe9c55238fcd92f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 2946, 33026, 2946, 2946, 2946, 2946, 2946, 2946, 2946], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1, "uuid": "e5a566f1-1454-524e-9b3a-439d3f77c2d3", "validation": {"axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04249559761956334, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e5a566f1-1454-524e-9b3a-439d3f77c2d3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 23, "MP": 516, "NP": 1527, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d4809ed698890ad82a3072f128d74c3820a9cf2c8c1ed44ab3d97cc6351c311e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2495ef88190ca36664b9556cc1978aba7f35b4afdb9a2dd8b4237e2113495484.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0f7ee7a4a5fb4a6bcd05f8b498a277af3123eb2acdf0bd715bb6d6262bbced63.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/df756443a3921e425e5e878a3bfda9c51af5b8c5b9d8e11a934504abaf31c802.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 33027, 2947, 2947, 2947, 2947, 2947, 2947, 2947], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "dc530b21-9ebb-5ee3-b81f-913961e55b86", "validation": {"axes": {"B": 23, "MP": 516, "NP": 1527, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006693470757454634, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dc530b21-9ebb-5ee3-b81f-913961e55b86"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/aecf6d2bac44fc607c769c6e75155433cf865b47ff4270b3054d136926c8fd6b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b1b5b7846f1dab5cd1d091230424d61edff6c8065ac23190dfe9c55238fcd92f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 2947, 33027, 2947, 2947, 2947, 2947, 2947, 2947, 2947], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 2, "uuid": "4f78d4b9-04af-5b89-b543-d646255f16a8", "validation": {"axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.04146892111748457, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4f78d4b9-04af-5b89-b543-d646255f16a8"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 23, "MP": 517, "NP": 1550, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6197f3ea3a562f4e0c1bfd66860277eb309f3e882d4f8ad78ed8e9ad12f17a7e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e46a7f38aa0257560f0b8174e9e305f50655c88aec220f7661cc68bd9c726ddb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b1b5b7846f1dab5cd1d091230424d61edff6c8065ac23190dfe9c55238fcd92f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cb9d100b30dcd38840e5972c2f43348729a0a654f3d311f1b405984802eb9885.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 33028, 2948, 2948, 2948, 2948, 2948, 2948, 2948], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "9e0e33c1-b008-5eb8-af3d-229a72ad0113", "validation": {"axes": {"B": 23, "MP": 517, "NP": 1550, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0069611696526408195, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9e0e33c1-b008-5eb8-af3d-229a72ad0113"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/62821a76a630729207c3b54b7056e094c15934e5552829ab000cb0915ece4474.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b1b5b7846f1dab5cd1d091230424d61edff6c8065ac23190dfe9c55238fcd92f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 2948, 33028, 2948, 2948, 2948, 2948, 2948, 2948, 2948], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 3, "uuid": "c925c9d7-8241-573c-87fb-6170953c8a40", "validation": {"axes": {"B": 23, "MP": 517, "NP": 1550, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.041415362153202295, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c925c9d7-8241-573c-87fb-6170953c8a40"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/de085bea68a1e94df52e7d840da4e964544f8745ee86ea5ee03ebc53806eeffb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9aa330675627db9704063bb387347b5ae7fbe867623a2d250c634de4574b5d93.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f7dcb24cb0737c7f92b54daf9b63c39cea601c8edba567516cfbe4bd911aa3c6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 351, "uuid": "121fa2eb-98f9-53ea-a378-2deac9279b94", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006448616739362478, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "121fa2eb-98f9-53ea-a378-2deac9279b94"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0d4468ab2f20fe3620880196b026f3662d97c867da9c8ae416baed7206b4974f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296, 3296], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 351, "uuid": "e7611e44-3849-5882-b0c1-e92cbd7f5dde", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03335760999470949, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e7611e44-3849-5882-b0c1-e92cbd7f5dde"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bbbf357b58963d5c8ea1d4da601354b745819f6bc962bb3eefccf980ce0593c5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b497d2b8b3c8aca38aa89047c1e165381df5f59dc616858c7220d8993e0f6ca1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/63021e1c04e1711256f6cd68502777b862dd199e43115c9d1a0e16d1d845b0da.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 352, "uuid": "01734d42-a355-5d3c-a16a-7027a4a899f6", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0064980280585587025, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "01734d42-a355-5d3c-a16a-7027a4a899f6"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/df218ba2d8898e0cdc566f6504cd21ca341b926fcbfe5bb671e90b45105c8255.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297, 3297], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 352, "uuid": "864c6857-c29f-53fd-a8f1-b81501f442b3", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.03316539293155074, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "864c6857-c29f-53fd-a8f1-b81501f442b3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8636a040f7f858fc2f03bdb33ee0489a4fedeeeef6e2878c0a8be1ef5c6476c0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b751607e293b80c4084fabd746037748d22aec0581089248ff06fe853c55865a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0fe287e9e6242b5ba0fc04a2c172dbf9ba3f69b1c5937a8e7a0554ee7b1011be.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 353, "uuid": "2518daf6-e652-5f93-8622-739799148843", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006214004009962082, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2518daf6-e652-5f93-8622-739799148843"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0579f851ba5e41801763d71da756aae81b9506a7f924b5c33184afd1048e6463.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298, 3298], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 353, "uuid": "fa60317a-749e-5495-a3f3-f9620c0cbeac", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03322890307754278, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fa60317a-749e-5495-a3f3-f9620c0cbeac"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/051f561c0ed570a2b9c22e3942a5f6bf94a12731d080b03a1463b11844ab7027.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a5ff07faf5c8edac53b71678b5bf1a284eca4a9bdf3ab8bec10d3692b06f1be7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b1eec09cfb9b44eba6ae5caf3f898a086dabea7e12655ffc4bb079a92e7e8ac8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 354, "uuid": "94bb564c-8763-5e97-8c45-6845cac03fc7", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006237378343939781, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "94bb564c-8763-5e97-8c45-6845cac03fc7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7567c6bd5c670e625af28c565f6bb841ae26afc679f5c915b3433494bc806091.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/87aa4e1d138880118ba51f6d159c683250a5971921934df9d29e8fce944403e8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299, 3299], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 354, "uuid": "4f9b7cac-e45f-537c-bc0d-7ece14e5f4ba", "validation": {"axes": {"B": 22, "MP": 52, "NP": 1143, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03397112479433417, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4f9b7cac-e45f-537c-bc0d-7ece14e5f4ba"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e489ccfcdd0ce7023031dc28208d0d04d6e73bd313298447a991f802d148a0f0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/dad7f4bc8029312356c4ddac932e6c45f5e937cef8842bd54d40b1ae82dc8aa0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d201c5318b998c218cf6cd15a3a4615b6799dc94ee61371000e4ecacda4a943b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 754, "uuid": "57662255-2933-5dd2-8336-64aa6018c0e9", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006210855208337307, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "57662255-2933-5dd2-8336-64aa6018c0e9"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f267b36450e33d3d362cc2a4bf0b25ed754b509b914d3db17a193c3112aff6ad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699, 3699], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 754, "uuid": "ab3b56e8-1497-542b-98c5-0f565d522c1d", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03436346398666501, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ab3b56e8-1497-542b-98c5-0f565d522c1d"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4c9d58c8b33bc2b8ef10e83c01c2754f9363f60f934dfa35db1b5f7326e46eb3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d8dadf7200af26ae60ab855f998f2d7ea2d0fa614ff74e3dd38b4ad6070e226e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e4199dee3958c8625a50491597d6e2fbd6ef9f1377e9da3dd850b6022193978f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 755, "uuid": "51f523de-e324-5c58-adc9-d1639d162545", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006215694826096296, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "51f523de-e324-5c58-adc9-d1639d162545"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/810b9cfbd5e5a9a0ce6cc83066555cb3c26a9d2eccf7b09b56e40bb190a24084.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700, 3700], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 755, "uuid": "8389c892-bf5e-5395-bcda-debaa3839fe0", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0341491480357945, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8389c892-bf5e-5395-bcda-debaa3839fe0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0a4509d9cb96317a6d272fcb2f9dd619c8a912431e0a58702d0c87f4577611db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b0f3a1db1242f419085877cafe0142d347b7da66ab54ad56ab033c07bfdfa9a1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3684f169ec37bebff764b8a263711bfbadb363ce2a452c141ae9f89e026d5572.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 756, "uuid": "3c00b71c-d78b-54f3-b112-9e22c30fab26", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006245373748242855, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3c00b71c-d78b-54f3-b112-9e22c30fab26"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3609fc800687b3ab46a44290d5dfa54b66fd07949f141a7e06800764550b1686.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701, 3701], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 756, "uuid": "bd420c7e-d7eb-55cd-9bdd-61590ba2ad6d", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03430905472487211, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd420c7e-d7eb-55cd-9bdd-61590ba2ad6d"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d6bd29c013578aa53c443e61e5f6d174c1abe8b9d61bc5af9edd29aacdbf4622.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8fe39e88f0e591bf840edc2a84920576d601cea58e81f7a5183955b872461f0b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a65e87ecb8998b59098c7787971843aea9e038c5d87568ccc5d54ebac7173597.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 757, "uuid": "18d9375c-863f-58f7-9ed9-318748343295", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006067672744393349, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "18d9375c-863f-58f7-9ed9-318748343295"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/73115e58376313aee5802d9d3a4f186fa5f93202f0a402bf717ac2143dab90dc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/15daec43ef933276b1088d6d1838bfcb80bef1ea0b8e5fb224acaa8f0c52ac61.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702, 3702], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 757, "uuid": "8c8097ac-9306-5e79-b95b-191532f0ac56", "validation": {"axes": {"B": 21, "MP": 58, "NP": 1217, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.03457778599113226, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8c8097ac-9306-5e79-b95b-191532f0ac56"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1e2c584d3e884a178cef8d481c91b16d8894c6fab30a881f2a781c042899de1d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/474bfad273b377a4a34136f877115200295d874af185d2272cc4133894568e9f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/35eaa02b81b266090022a5e684068eab6445d925ba084857ed3118aa06ef6194.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 899, "uuid": "a8efa001-67bb-5c58-80f7-02664623b588", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0060572512447834015, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a8efa001-67bb-5c58-80f7-02664623b588"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/eb0f7ac1a0c17a5f381cd1cec8db72179a9961d2bea156b1027e449b4ba87046.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844, 3844], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 899, "uuid": "02ee0c99-f587-5ced-b116-7b58cd8559a5", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.033648169599473476, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "02ee0c99-f587-5ced-b116-7b58cd8559a5"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6492379ef1aa0a676533ae930b879d7e25946b7cf9c6637efaed5b5eaaf41609.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e06351c7b8916ade56eb8989e6030d621276f706495852deb8393c120bd4e38f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5812ebce045e9a9091520b153dfe7cf1d2aefad52ae1d2492f1bdf6aaa2920aa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 900, "uuid": "7a4fdb22-2165-5d67-bb42-29846288bd9a", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.006133459974080324, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7a4fdb22-2165-5d67-bb42-29846288bd9a"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/384340b368e521229becc472ea06fd06444d6f0efd331bac9963626424097233.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845, 3845], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 900, "uuid": "6bc8d826-0b06-5fee-b159-77fd082220a1", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03410201473161578, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6bc8d826-0b06-5fee-b159-77fd082220a1"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3d4d0198ada34bb6b3a89131e6536998e59b04523d3618cf0475fa508bd493df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f01de318bf4e0f2949e2135351525cec1f2847a9b3cc6ba96d07c38432b2a477.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/58e2b348b5bf8b53f9f5f0cc3ca27a24cb155f12a837936ff7da842f723bf999.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 901, "uuid": "af51e75a-f41b-5bf5-96dc-f210165a3a99", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005838514305651188, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "af51e75a-f41b-5bf5-96dc-f210165a3a99"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/39c80239595f2b67a2a4604b2b7373233e0ffe7518a9f706eb7054ecaf12c0f6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846, 3846], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 901, "uuid": "ed21efd0-f769-5ef3-a587-974d774f39a3", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03680899599567056, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ed21efd0-f769-5ef3-a587-974d774f39a3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c84ab2245b94a3c1538cf2675ddda689c832df15d19a09f691769344f6fee8b4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0c31006f90cc25ce19b5c02eed260d73cfd609fa862039a44589dccc45cd2635.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e42ff588dde8e3c0793658cc7aac0f91ce98a9d33940af1d06c3cceade3bfd24.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 902, "uuid": "232e9d38-40f4-5011-88d7-610037c52116", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005822970997542143, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "232e9d38-40f4-5011-88d7-610037c52116"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6f3426ed07a9a5dde917e3bd8f86a22b1bdeac08032073e615d4e0801d6562fe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/717b9b16497bcdb0b8cf400e260eb2116574477849558067dfcfdfc93b7d964f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847, 3847], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 902, "uuid": "74eae518-bb9d-5242-9873-727c4818706a", "validation": {"axes": {"B": 20, "MP": 61, "NP": 1219, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.03422632487490773, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "74eae518-bb9d-5242-9873-727c4818706a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8ed900936b3819396189727e9787f0c70ff390cf0573e6f6c8a5237c36cfe7d3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7a46f7ca47338f4fa57a6d8fdb800197774b3ed20af799c6403bf9683fefde26.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2b3302e6a6c84b9f964bb5d108d3f2aa0642e85aa5c8e66a332de95098dfae8c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 935, "uuid": "f0bd71f8-3acc-537d-9d9a-c05bcafbacd1", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005778278224170208, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f0bd71f8-3acc-537d-9d9a-c05bcafbacd1"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5f61ad0cd813a1088bba75f95d07319d8669360a21084d4481a4b79576314508.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880, 3880], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 935, "uuid": "6ece5649-2c19-550e-94a4-7cf60234a5ca", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.031692517921328545, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6ece5649-2c19-550e-94a4-7cf60234a5ca"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/06cc1f356c5b69ebb97cc1629060fa6b62b2260a9935b307a1678ff5378269f8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2ce67301cc5465cf15212c038ddb3b7cdf7886044d6a24f524abfd6c535a8b25.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f34b40b00b63381ffc1a5fe7378e710ceb57915df0463395603d5ec51e1032f8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 936, "uuid": "1a9ecd12-c4af-514f-864a-147776870ec4", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005763749126344919, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1a9ecd12-c4af-514f-864a-147776870ec4"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/dac8c308fcfc85349ab9cb8b69c26bf16c1962180934f5ab9662d8d1da5e1c75.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881, 3881], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 936, "uuid": "442b709e-d737-501e-a585-3df54fbdd195", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03186331735923886, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "442b709e-d737-501e-a585-3df54fbdd195"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a155b95f2a7908722ec0474f16f2fddd53b4ff8478ad1858aac871b33bb86d4c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/793a7bb500e1c9f5fff7c70d83cb3831083b3b43bf1fe27c73fd8200abac3129.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6dd38ac38702024112b20b0435c56c3e91db18187e9072a99fb228984f4a8c8c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 937, "uuid": "1018ebf9-56ee-5271-82a7-f78e201b8ebb", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005550278350710869, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1018ebf9-56ee-5271-82a7-f78e201b8ebb"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a86e208f1f67c39cb08b8e9944e544c6b9709c0d8fbbfb28002602859ecc3903.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882, 3882], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 937, "uuid": "92822b41-eb57-5549-a45d-366047e42f6e", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03178327810019255, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "92822b41-eb57-5549-a45d-366047e42f6e"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f16368c87596a00d14b53a6d4db3594113f639892318382d35d1e76e5e8bf3e5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a4a9ac274fc8776281a81c3ada46b09bd004b9deadfde4ce7576f092b07d45f0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fae231e57f85d2d3347c0928a35803ac8aa915055145e6fc53d5de4c177929cf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 938, "uuid": "3a22b224-a255-5fe5-9d48-201b7b41e8d8", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005545355845242739, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3a22b224-a255-5fe5-9d48-201b7b41e8d8"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f5c69cbf453b315ec328e7f2172024d6365455196f0fc0a792e4b878be6394a0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1e4d3c0935ff1313e27747ab22c924a75d2d7a110328163848e3f22c75c400df.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883, 3883], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 938, "uuid": "ef725213-e67e-5af0-bc6e-29b3944aafc0", "validation": {"axes": {"B": 19, "MP": 61, "NP": 1158, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03227614797651768, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ef725213-e67e-5af0-bc6e-29b3944aafc0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/72e8eb46d0423ffd7687efb01f9683b92402641c220b125ae3124f987985a022.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7918d27d0933e12f4c4bca7bd0f1e9914f76d83626e982e35e6d22825f716da1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cccf7e4110ee1a834b7010489e4e70e09f75f165e280da14d96aa92cfa5c7208.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 954, "uuid": "04482113-d660-56dd-9478-dbc0f050fd02", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0055061341263353825, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "04482113-d660-56dd-9478-dbc0f050fd02"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f2a9c3d4c9b4b23039cef34644aa820c2925d6366d783d1f464b3c415829a5fe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899, 3899], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 954, "uuid": "083e42e0-3128-5aa5-9911-8bb517cdc688", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0304143400862813, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "083e42e0-3128-5aa5-9911-8bb517cdc688"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8e31db1ea60c94949e5ee170893226a4690ff02e34ed0c414198b33950bba71d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/430f5594b29abacecfab47cee9e09f55da83c9685158c2953fc91effa4a6ac8c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0375362676db8fd19cd880a0ce8a69830ab311a4fb86a542a032e0611e9e30be.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 955, "uuid": "bc654534-c52c-5ecb-a041-2d1a27ce3ca6", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005518301855772734, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bc654534-c52c-5ecb-a041-2d1a27ce3ca6"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6c90ebe244840e5050afa8708afd55f1caf7eac47b27d793853a0daf6457451c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900, 3900], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 955, "uuid": "97d8da85-d282-521c-8e92-96a5dd2340e2", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.030403156764805317, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "97d8da85-d282-521c-8e92-96a5dd2340e2"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/03f756da5e09cc6c31f35244c4447c43e419ae5f16734801e3ae6cea1b592997.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5a682d0c4417ed6459b78b6a6c82d866058bc367da074105554ad684c19e2c6c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/72cefdcf69ed5aedca7f8593c903b72c98d91ac5bf4c79fe12b13d9202c46aaa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 956, "uuid": "caab1a5f-714a-57a1-81e1-a6c119fdba22", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005573482718318701, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "caab1a5f-714a-57a1-81e1-a6c119fdba22"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fc028c6b23a92e9478957ae570ab50466288c42bdcf240ad0481cf7db5b18ccd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901, 3901], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 956, "uuid": "b4384cb8-c729-524f-88e7-85e7aab4793f", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.030967104714363813, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b4384cb8-c729-524f-88e7-85e7aab4793f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b4927cc2fa63420372c9ed4521da0745cce90a2fec3b95b039d27eb38370cb2b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/95b43518d2b10d3bf51341824ccabab7a7b45a4323a9c9f124716db0dec3d1a2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2239bb568b8e70498b1b24854ef836bedf4e08583f353748026021f37dd575fb.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 957, "uuid": "bcb4b8aa-6de3-54d7-96cf-35ccf6cd1022", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005315464921295643, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bcb4b8aa-6de3-54d7-96cf-35ccf6cd1022"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7c0532d59c7ff5cd9bc539a239cd050ef1cac917cdb24520200e6668e1891ec6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/faf93d327cdf51b1e0a615d822b9a4fa980aa0dcc62b49af7156307f7fd06b4a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902, 3902], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 957, "uuid": "bcffef9a-eb8e-5e4d-95a6-6a0a65e55038", "validation": {"axes": {"B": 18, "MP": 61, "NP": 1097, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03021505381911993, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bcffef9a-eb8e-5e4d-95a6-6a0a65e55038"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 17, "MP": 63, "NP": 1070, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/60c0de7fd1134dfd1f1153a5950d3fd9cb76f158c501c315d6da03c2f00f8747.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9e8a6d260268cf271c26997c02c6c3e6ace61a6561f86c698bbdd639ba994765.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cc7550a60a3ed58861af674466a562213ec780c119ac67d0170c5ec028e75584.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/53431a011a505c98d1dd5bd9f131274e201cebbdf5434eef753a6607a4efd1b6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1055, "uuid": "9fa38acd-abc4-574a-963a-6b59e0e31913", "validation": {"axes": {"B": 17, "MP": 63, "NP": 1070, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005304791033267975, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9fa38acd-abc4-574a-963a-6b59e0e31913"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 17, "MP": 63, "NP": 1070, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/3ad1319dcc42a8d66d293395bfba28f2a428a3b71986b1b6fedff7d1cb71da69.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/cc7550a60a3ed58861af674466a562213ec780c119ac67d0170c5ec028e75584.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000, 4000], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1055, "uuid": "e1567a4d-7a5e-5e6e-8a58-9a58a5aa3468", "validation": {"axes": {"B": 17, "MP": 63, "NP": 1070, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.029194852337241173, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e1567a4d-7a5e-5e6e-8a58-9a58a5aa3468"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 17, "MP": 63, "NP": 1070, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/310b3a0cc0411f8cc57f3ef3f75a6072f881b775c53bea5d45513eb922de1492.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c4ec759088e2542f97ce1fe34e48e2b86b58af3ea1355455845156d710d80dd6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cc7550a60a3ed58861af674466a562213ec780c119ac67d0170c5ec028e75584.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9822aa0f5041eafe110b8ca344d2ee33c97530bdff4e351b15b8f0e0aec2f143.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1056, "uuid": "8277fb06-67ca-5af3-b956-50ee84f22cb8", "validation": {"axes": {"B": 17, "MP": 63, "NP": 1070, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005309311207383871, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8277fb06-67ca-5af3-b956-50ee84f22cb8"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 17, "MP": 63, "NP": 1070, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/913d8cc1fc47824553338d974604c347d5326cca6aab903c1ae5ffd5b7d33d7f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/cc7550a60a3ed58861af674466a562213ec780c119ac67d0170c5ec028e75584.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001, 4001], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1056, "uuid": "95b781b3-603f-51ee-88c5-413a63faebce", "validation": {"axes": {"B": 17, "MP": 63, "NP": 1070, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02979897102341056, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "95b781b3-603f-51ee-88c5-413a63faebce"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3dd5ed422beddfa9e503a58cf96c218c7e17d6162c2425ebe83f1e7434042d95.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/26df747d9d397fc44b4d7d2fea5304fa41f64631aacf2726cb2e6da473b87e14.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bd32ebc3055a2a8f3e129104f536d7c4d4a7c7c3eae53e0e65d3a21b582aecae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1057, "uuid": "ed7d394f-ec0b-55a6-93ae-c8d88430ec2d", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005196934100240469, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ed7d394f-ec0b-55a6-93ae-c8d88430ec2d"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a0e6f3af80ece982fa5fcef83c18bba0df131913eab9f4111a4f67d17f482aac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002, 4002], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1057, "uuid": "da9612a7-a4c0-5008-b65c-d4cefc82c0fd", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.03166577219963074, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "da9612a7-a4c0-5008-b65c-d4cefc82c0fd"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9a66475cc56b5e28491db6292858f7c95f6e52844340cc24499317926c0170c6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ca807017aba51f1e98ec076864ee2aab6857551e533decccc15764d6806ca0c3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/40b7fc894850ab7f3927e4d3085a572e4c1785f22a6a89b6b055f5f22f4ec1e1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1058, "uuid": "61f52c3c-66fc-5681-aac9-441e37cffd55", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0051085022278130054, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "61f52c3c-66fc-5681-aac9-441e37cffd55"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5dffc729f016dc8439a20d754b32b4f45549304f369c1cf8cde61e42c368c461.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003, 4003], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1058, "uuid": "ee88b676-3b09-51cd-b734-01d5a1b9d34c", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.03125], "seconds": 0.027891026809811592, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ee88b676-3b09-51cd-b734-01d5a1b9d34c"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fc23188bc109ae3b73f3ba79e70a3ded9d4e93aa8b68a0ee64e310e4fa62754c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2f412f4175cb63fac0837bc8856b5ed5815a8d4dfe35e9c75de5ae10762ad5e7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/80ff4000de705215e09f7c85407a450ea2240a52e82e7e37bb7b2831b47e4bdc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1059, "uuid": "6fc424c7-18d6-5838-98fa-69570862a599", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.00509605323895812, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6fc424c7-18d6-5838-98fa-69570862a599"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/dc30839b425ca17464682e4fb5181ea25d63e6620d28f5dfe081ec35d89c9630.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004, 4004], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1059, "uuid": "11c1d515-fca9-5920-9717-be39327c06b7", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02757177595049143, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "11c1d515-fca9-5920-9717-be39327c06b7"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ecd6636aae2152c1c2165c4ac4f22f65a1780ef3d494f50b6f20e1d262561d42.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/89fa58d11ffd01660217d53137155d5440e742c6cf3ab6d31e10b2815a8f58e2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/725e6e263a082c8f79791121521444ea71a4d5d283c019c2d2c9111d45d19f0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1060, "uuid": "04db0d95-e1f0-5b31-a3ab-4a72f2f290c7", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.005062700714915991, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "04db0d95-e1f0-5b31-a3ab-4a72f2f290c7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/31b969fdd8d4a921b3a213cf6c43f7d24c5aee38a2a53052367f942ed9186403.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9612457384c1885a1ce1fc15f37b33c091934c8669d91fefc82594da9ece1d2c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005, 4005], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1060, "uuid": "28d11f38-45b5-5f03-94a6-2d4da1174e3a", "validation": {"axes": {"B": 16, "MP": 63, "NP": 1007, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.027850419282913208, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "28d11f38-45b5-5f03-94a6-2d4da1174e3a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3455ac219714bc9927bbd795faeb3eda6cb1b724bb4477d1981a6b839d3eb8f1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d5278bf331808ab2470f3bee8925aa3cc10d919faf95aa272206b57ce9d79ed7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c955f314f15c4d6e57cf0fc9ce8352ffbe3304cbcace2b835482f2709b5be9c6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1068, "uuid": "54660deb-e53f-5695-bde3-e0036e163b2b", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0048505449667572975, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "54660deb-e53f-5695-bde3-e0036e163b2b"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a50f1f79e553c3ffee562303f9ea325569a93641aa3067d533184ab271429882.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013, 4013], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1068, "uuid": "e13d97a5-cb78-53cf-b556-24b20c61237a", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.025956761091947556, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e13d97a5-cb78-53cf-b556-24b20c61237a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c84257356c8490c273a779ad8fcf5ceae8bc5c6b2d06dd0ac70b56c56d718886.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/36be477ca379ef71dabb23e5a0c57e733026a6782dcdfa667110c0336881ccc4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/087dbc76abb070897006c18bb665ffe0981864237b05859dd55e6eb6df041100.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1069, "uuid": "28a0babe-dae2-5abe-8b6a-09571fa1f8e2", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004851930774748325, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "28a0babe-dae2-5abe-8b6a-09571fa1f8e2"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/27285cf9ee4c91a43b6f5cb0d82be50ac169d1470a1f4c9721abe983990596f1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014, 4014], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1069, "uuid": "9518bd14-8842-514c-82f6-28dc7207b025", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02603803062811494, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9518bd14-8842-514c-82f6-28dc7207b025"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7a32f58da83872d6bfe1d5642415779ea7c92819cca50c501478eef03db8d2f5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/609df8e7b146235077e5d7c0e976c65d5879eaa5398bcaaec65b610145f9d6f4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d63b50ef51f97934144dab841b2c01569a4b1d7febde4fe9afca34aebfbdf6a0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1070, "uuid": "b4a18b11-a27b-5944-adf7-1695c3fd1536", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004787751007825136, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b4a18b11-a27b-5944-adf7-1695c3fd1536"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8fa5ea1005d2b79eb7c5a079022b8110572f5afeda4ca01111f4f7f27e3310db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015, 4015], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1070, "uuid": "8efbb4bb-4a11-5b62-b757-ee68269bbabd", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02597333211451769, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8efbb4bb-4a11-5b62-b757-ee68269bbabd"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4ea0185d7ba5457fc9bd0025442460566671992df1e852a99862ec32e14b9130.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/058fe63657516008c9d55b044c2a4040fafd15c03809b367e1d046a818302a85.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6811d9e0b5d6de194ab91280c54f22f2f0c925d97138558796e396feda99809d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1071, "uuid": "82180535-bba0-5fa2-8a24-f7f81db095b3", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004816650878638029, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "82180535-bba0-5fa2-8a24-f7f81db095b3"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/49516fb3cbe98dafa7b6714ebc5d09e65ffd09ed658065e562e0759afa71a73b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e76ae94784d650299544d889ee801395cc91135e1267b8b867a0f8e0d5aac119.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016, 4016], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1071, "uuid": "33958fe9-f003-535b-b2bc-a1b1339ce749", "validation": {"axes": {"B": 15, "MP": 63, "NP": 944, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02595904003828764, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "33958fe9-f003-535b-b2bc-a1b1339ce749"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6b186e5d1020f7a0e3e01f86938af5b888c3cabbfa458949dd3d2c23846bafa0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/739020acd2b6347f799ebc37f2bb56983b6ba8c6e6066da944ebc08e27c67b6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4a1aa7db5343be942a02335e6325c621d31516532338787ac8c0c8d4bc4bb59d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1176, "uuid": "35420a6b-26a1-5d8d-94f0-51050320239f", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.00455931993201375, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "35420a6b-26a1-5d8d-94f0-51050320239f"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b9d9f67fecfe7765b134c9dd0df3b3e7f483079a1f9fa85d7f789f55626da42a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121, 4121], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1176, "uuid": "c9cdaa2f-359b-5142-8823-935534fc37c9", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02474446315318346, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c9cdaa2f-359b-5142-8823-935534fc37c9"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/231c465f051c1cc9193efbddfa629da27474f2df29b7ad7752a39322acbede2d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fb7085e74c6c10a9c820291e3fe73841a6f18b8436972c47a5116630d30fbf7a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/43c72292a3f6b17eb4a93e4b863d4cda8e2d676b5831f5d86ecc4a991dafe622.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1177, "uuid": "171dc6df-f1ff-51cb-a2d2-4ef369ac2b13", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0045750997960567474, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "171dc6df-f1ff-51cb-a2d2-4ef369ac2b13"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/605f0b12b6ff43dcaea83c3284f9efb5ea9165489f28eb83fc512f4e35060110.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122, 4122], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1177, "uuid": "8d9596c5-4160-572a-9542-3ad4c242bd06", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.024800621904432774, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "8d9596c5-4160-572a-9542-3ad4c242bd06"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/884fa35e7e4875dd6fddd39e4a964ee8e4599ec5e7a60669c61dd278d23743dd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/979ca71ce2120abfcbe80cd4b85dafdfa7d9887dc9b2b44902f7abf8f1f9c0e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/925fc23ac5bd4ac50950f5127fdaff1284f61e25a3309d89818eb9e57ca680e5.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1178, "uuid": "aa1ce94a-e70c-50c0-9f74-3e2cee278f77", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004444255027920008, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aa1ce94a-e70c-50c0-9f74-3e2cee278f77"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f395d1f29ad335a04063065921dc4b2e1052237f5377955ec49c9076051e93c9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123, 4123], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1178, "uuid": "60651d55-64e6-5b0d-bbd9-5395c05ba146", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.024785593152046204, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "60651d55-64e6-5b0d-bbd9-5395c05ba146"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/dc81fe66cf8fa14ada6bc98be55e84a9a3eab8d517d6264f474667c0c7bc945c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0e3c9e309c4fe1e697b2907035d382df8f30daf790b2e2281b4512f279b507ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fc78f36d912e39298d173abb869af563da72b95a942dd768ebb6ff80159e8ffe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1179, "uuid": "49a6bfed-83e4-5740-8a15-d4febe3acd45", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004559503868222237, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "49a6bfed-83e4-5740-8a15-d4febe3acd45"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2b8eaf5aac14162d92168f850f9a4a3a1319ffaf66e9fe07e91fb1aa1b404028.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e5d1467cb988bcdc785f9a9387e0a404a95bac2e0a0b741837be0a168d2b787c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124, 4124], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1179, "uuid": "c30fb578-448c-5e7c-8652-4aaf27fa84af", "validation": {"axes": {"B": 14, "MP": 65, "NP": 909, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02480331715196371, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c30fb578-448c-5e7c-8652-4aaf27fa84af"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/79fffcb273e155effb280dcd9a9d0df963661bfb4cf6d39d5a11b98ffd35891c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/aa801b2787f0637d3686cab25865463e42e98c9ace929bac0aebfe3c53c02787.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/68ea03a5849e531339fc57c7ba7aa15b11528edd3a757649de86883b248a510d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1184, "uuid": "a646721d-c7e4-5c06-a4cf-442db0513825", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004297700710594654, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a646721d-c7e4-5c06-a4cf-442db0513825"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/487234e6dc6995284b3356381c73ca246dc035003ff699068d0b3be8a8e5a34e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129, 4129], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1184, "uuid": "356b235a-a39e-5510-ad50-1113368af867", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.023279639892280102, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "356b235a-a39e-5510-ad50-1113368af867"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/080cdf5ad64aae7944435e71a000927ee11deab9f44712acd2beb7d179c829a9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e6b150ec5b2f7719644f4f3812a472b4b8982b75e97be516e08171e6af6e050f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1b44b16ea79929165074b45fd3cca2efdefe2c3969ae34d7877c25c4347dc632.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1185, "uuid": "461a7d89-d279-5d75-8647-a44442709c65", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004295635037124157, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "461a7d89-d279-5d75-8647-a44442709c65"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/76612f9b3ad3c2c3a5df001145fc5e2e2ddf5e265d00921a1a5f024e5acee5c4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130, 4130], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1185, "uuid": "ca28a154-4858-5563-875b-6d4b983b3f1f", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.023234956432133913, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ca28a154-4858-5563-875b-6d4b983b3f1f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2f89b51fbb5688fee18d357bfc9f82b44444cfdc47bb0a14ffc14b5c71637724.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/35d1dab6b7d4163f51e63bcf33af69603473f21c52257e8e4a5e8b641247682b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9fac8b38972cc5d1f9878e4d3c5919dbde041c481099d21d66d9c7e549e37656.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1186, "uuid": "3ec575f1-d536-5e37-94f7-e42b312bcabb", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004131546709686518, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3ec575f1-d536-5e37-94f7-e42b312bcabb"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4e05bc3828b94e402187193175117876e5f52d2e8cf5f9873c486284c95cce14.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131, 4131], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1186, "uuid": "7f201a16-99c9-58a7-93f1-667b9478c166", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.023138949647545815, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7f201a16-99c9-58a7-93f1-667b9478c166"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ac358882a01d83035bfff18154f2c840a4c087ef64765bb7e704a5ebb2867653.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e29ea76899bfd7de8eceeed4f1975fc5442ac765525bf84aba0fd207116adb6c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d112df72d9911aa47f9063c70b25f909a3ea4bc991a15613347ec0482035eae3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1187, "uuid": "2878bef0-4bc3-5841-b9a5-d44a31f23f98", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004343045875430107, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "2878bef0-4bc3-5841-b9a5-d44a31f23f98"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d888f5c65a32d4ea8d3c4e22fc55b5f6454eea0c1ef0ff7abd67f69bb3a35012.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5697e9bce1e94d139787fa9574317374a9382515c522f041eaf2578cbb289918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132, 4132], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1187, "uuid": "79e751f6-0044-52f0-86ae-8f6e7720b5ce", "validation": {"axes": {"B": 13, "MP": 65, "NP": 844, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02311041997745633, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "79e751f6-0044-52f0-86ae-8f6e7720b5ce"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8c0dbeed0462399ada66f459b03d65b60d41900c13c1de06bf7386c3a61c7043.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5f006bd50aa9c9251f62b2c922e34bad3b9fe124f4b8e421edc08644c796bd25.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6fcfd37d6d2297180a776851f581c932b89de001f181232e406cd571b7bbfc6d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1243, "uuid": "9ca68031-6b0d-523b-bfbb-b6afba51a54a", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0040741306729614735, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9ca68031-6b0d-523b-bfbb-b6afba51a54a"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/894ff6fbb185894f9ce7615dcc0eb6a2aa9ee329ce489b47498548040980b833.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188, 4188], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1243, "uuid": "90838d68-26d2-5ed1-9091-45fbc95984f9", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.021693270653486252, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "90838d68-26d2-5ed1-9091-45fbc95984f9"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ab9f625594ffd68e2b0defa7fc8deb7575ce2f8657efc257049814a37827a45f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a304e901ed8960c740a67be85bee3c1258a018e9290a3367407427e306a060c6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b411163b5a293938e5fcaa8918b00d0431d4b8b711637e5780f1081334ec28ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1244, "uuid": "5ab9a539-5bfd-547f-a367-17fd758f5858", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.004070143215358257, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5ab9a539-5bfd-547f-a367-17fd758f5858"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8c9d5167ddc6c375067ad4cd48765dda4775b4d4c09dbffe5c0d0bdb0a892918.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189, 4189], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1244, "uuid": "254a3fd5-6906-5874-9317-aa68d52c036f", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.021799088921397924, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "254a3fd5-6906-5874-9317-aa68d52c036f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/afc7cd66acfa6c7a6a37313dac6e6affbeb516f6637fed725c860c1a8216019d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ad3f5650ff5e30da9c1d2cabd6a337697e2c03f07c0e23325e48b4dd7ea8946e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d7744bf4aba913ec7551ce5cd9ad3af3c0a938003bcc272b06b3fb63cb464ad1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1245, "uuid": "e690a1a1-ee7c-5b6e-9803-4d57a69febd2", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003916776739060879, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e690a1a1-ee7c-5b6e-9803-4d57a69febd2"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d3c2b812ca9ad333fe6a7322106d329b54a7204c6f261ca9f58610331aa87daf.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190, 4190], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1245, "uuid": "50dab7fe-d7a2-5968-9360-bcf785e955e5", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02172673586755991, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "50dab7fe-d7a2-5968-9360-bcf785e955e5"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/37efcfca416656fd12a08a5d14a315af6390533b7cc54e2bf8a95d849707eb2e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/26ef2a97c120494e043dd53b3beb7febd1b9c9839851e02549df1bac8bcb545c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5b410fbb0e3010549d6901f5c7b20dfdc8c522e6949192b07099de95c90c75ad.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1246, "uuid": "0bdc77b8-8a4d-5a8a-baa6-a348a2e387f9", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003921364899724722, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "0bdc77b8-8a4d-5a8a-baa6-a348a2e387f9"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4b3befba026d286e49211f0a901d3094d01a1a2f326cf599581034598c25d105.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c0518a0e60331ffb4567070f99397a8655f2ee2f6ce468e2f7c2170e116733ae.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191, 4191], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1246, "uuid": "4d276b6c-6348-58a0-b030-48b4e5a1c284", "validation": {"axes": {"B": 12, "MP": 66, "NP": 791, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.021704677026718855, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4d276b6c-6348-58a0-b030-48b4e5a1c284"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/22b8f0b937b107623dfb5808aa75e2e514d715ae5222eab72972f6b3d4dc960a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cd67bf318c64fbb1504f152d6ddaff111f6ca0d11b83ee7c8b4df7ede913773f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f6b12013a98f46dc9bb89069176208a95056ce930c80a0563666f9b8af10c645.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1249, "uuid": "abeaab28-d2e5-5ffa-8d42-7f4b029f3712", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003824952058494091, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "abeaab28-d2e5-5ffa-8d42-7f4b029f3712"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/04191358b4e0c5983d0be31850f9a80a3bcf2d2a582f2ead62c97ab1d7c82957.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194, 4194], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1249, "uuid": "fd0b6413-c7f9-52fd-8674-099a19ef8df1", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02000590367242694, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fd0b6413-c7f9-52fd-8674-099a19ef8df1"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/da4b2038b9a5793eab12d1b087f6ec041995af64a15c42d54df03ad08acc8a35.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c259286226e632d4d23ebfca4bbfc6dff400b8d8716c39ed943342b58c64764a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c886749629b891fa2a0b52d2aaead130f4350c058f10f49d22908e99f76cd6ac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1250, "uuid": "d508c822-02bc-591e-8187-b71842363950", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0038016908802092075, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d508c822-02bc-591e-8187-b71842363950"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/6e7b4a4bb96e5cf5c7e5a1478742fe4624d21b6515429e0e94615f867e47a92e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195, 4195], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1250, "uuid": "70fcde89-7f73-5c07-a65c-279c51c4e2fe", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.020212492905557156, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "70fcde89-7f73-5c07-a65c-279c51c4e2fe"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0128a97d24efcc1f32925b7bb355b7e903b4dec7a86f1e6fa3139d5d3a3eb40a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7942d0fad41c321016c16c0d3c1569bec82b65de79c6535cd402008579d9c6c1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f8d9a2e697461454dda492c3ec2eaafce2204e8694c435c7410a58dfe4aa61c6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1251, "uuid": "016cd5aa-8b47-529a-ab37-dd432df64b4e", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0037941080518066883, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "016cd5aa-8b47-529a-ab37-dd432df64b4e"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/690d09895dfb5f04be907be9a17dc319e176b33715d9089bb93e27e3ae5957f3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196, 4196], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1251, "uuid": "1f0d92a5-f9f3-5570-be8d-c3ebd628ca6a", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.02007369976490736, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1f0d92a5-f9f3-5570-be8d-c3ebd628ca6a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b5b2496a7ecaa5f1a9750a7d06f3edb9b14bdd76c7449f2bc98813175e492dca.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3ca18f2dc7cd9ce42b030c7d326bf13e47051a0c4b77677f2a7c00668d5d0674.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4e363a288be354e9f6e898dbe28b7c8c8037ec6a02682586df4c8405ed3ccd3f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1252, "uuid": "7e2cce7c-efa7-59b7-8657-efbb38aa323f", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003793440293520689, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7e2cce7c-efa7-59b7-8657-efbb38aa323f"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/15814fd0f77a7b167e7698c7918f5e6f9a74d0a48c235758087f9d633b6ad911.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5319c266f877b3fa16edb233551eac7b9d49af9d4ce595f423d3999cfcd0941b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197, 4197], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1252, "uuid": "7f64e56c-de13-5ec9-b8a4-30a3c308448b", "validation": {"axes": {"B": 11, "MP": 66, "NP": 725, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.020174923352897167, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7f64e56c-de13-5ec9-b8a4-30a3c308448b"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 66, "NP": 594, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/dd0869ca53cab408bee4e11044f431d11843c45078d4717a1c41dbb4312aa681.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a1479f7f16c902ffd1beedecac6d1308b2ecd6be610c1d8e3065073bc07fdea9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/29ed899b79af388f535677e3fd083a7d5bf54e0754813572f8a81aff15a531f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/13c6ee0c824505d38400ee234318b5a42c4fe70dbeba694b63aa63a3dd238b4c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4226, 4226, 4226, 4226, 4226, 4226, 4226, 4226, 4226], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1281, "uuid": "ab272092-081f-5879-816f-c4a03edc2b60", "validation": {"axes": {"B": 9, "MP": 66, "NP": 594, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.00331629766151309, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ab272092-081f-5879-816f-c4a03edc2b60"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/f60d9c83f41c475cd6ca572c3618317ca95c87f807cf3b330e5975fb002a4ede.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/86c03d0b6d4eeadff128875a598cb8d873f8eaa70a0946926fc815d69d3e710e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4226, 4226, 4226, 4226, 4226, 4226, 4226, 4226, 4226], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1281, "uuid": "bb8fc47e-6b21-5d81-800f-d3e249e6125d", "validation": {"axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.01678940514102578, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bb8fc47e-6b21-5d81-800f-d3e249e6125d"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 66, "NP": 594, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1b43e25b7152224bc33b42ed4e24087025b9aa09ad3c760a286745be9fe736e9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/12d4ae368e79f1c37a3146adcd89d3ab96ce4ed0509ce72b29c4e34cad721478.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/29ed899b79af388f535677e3fd083a7d5bf54e0754813572f8a81aff15a531f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e7dd881af6584ad63da4dc8bc097b541a1f8cbbfcfc746a5705bd74b8086594b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4227, 4227, 4227, 4227, 4227, 4227, 4227, 4227, 4227], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1282, "uuid": "bbbf2523-e1b0-5482-92af-fadac59575ed", "validation": {"axes": {"B": 9, "MP": 66, "NP": 594, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0032900008372962475, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bbbf2523-e1b0-5482-92af-fadac59575ed"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0f227396c5a17f90655a843bbcd5675a50d3b0f80f7a1094e3c17f0415f1579e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/86c03d0b6d4eeadff128875a598cb8d873f8eaa70a0946926fc815d69d3e710e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4227, 4227, 4227, 4227, 4227, 4227, 4227, 4227, 4227], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1282, "uuid": "33a372b9-4455-5891-b100-1ccda8e98c17", "validation": {"axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.016914537642151117, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "33a372b9-4455-5891-b100-1ccda8e98c17"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 67, "NP": 603, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/96a378a47e2868dd0f701a6ce57c405c1ec458992321152d749f3065662616e6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/08e62f7b1e851e892e78261600efade1d5cf8f960ab0795267b94aaff5472d6b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/86c03d0b6d4eeadff128875a598cb8d873f8eaa70a0946926fc815d69d3e710e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/68608b542e75b80583fd31f5dbc91e2f0264114a7276f9dd060d0987514493c4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4228, 4228, 4228, 4228, 4228, 4228, 4228, 4228, 4228], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1283, "uuid": "bd4fca7d-c68e-56d5-8d10-3ff4c55fbd84", "validation": {"axes": {"B": 9, "MP": 67, "NP": 603, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003296379931271076, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd4fca7d-c68e-56d5-8d10-3ff4c55fbd84"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7c31c4f4bb7f84d4b595d43c28c26b3775df4ff3de1a91a63f1e662b2c5b94c0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/86c03d0b6d4eeadff128875a598cb8d873f8eaa70a0946926fc815d69d3e710e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4228, 4228, 4228, 4228, 4228, 4228, 4228, 4228, 4228], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1283, "uuid": "d42f6445-22db-5f81-bc61-a880debbd6f0", "validation": {"axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.016782330814749002, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d42f6445-22db-5f81-bc61-a880debbd6f0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 9, "MP": 67, "NP": 603, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3a7487de0d58858b1eb5cbaf055de799f73600395fc318c9d6c46fcb8d834d17.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ecea209adfcdd64cd533aa2ebd5f62b4ae80d2ee74231e4fe77b012137357618.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/86c03d0b6d4eeadff128875a598cb8d873f8eaa70a0946926fc815d69d3e710e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/79e20b31bba3c4cc7548dafbc5b62aa6c0854a6335d575e360ed966e73a02e99.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4229, 4229, 4229, 4229, 4229, 4229, 4229, 4229, 4229], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1284, "uuid": "3bafb485-dac3-537b-8544-989add8cfe53", "validation": {"axes": {"B": 9, "MP": 67, "NP": 603, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0033061369322240353, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "3bafb485-dac3-537b-8544-989add8cfe53"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2c78228108e79650a3a016a80bcc5a7d0abb0356867430f3daccba305d3742f3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/86c03d0b6d4eeadff128875a598cb8d873f8eaa70a0946926fc815d69d3e710e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4229, 4229, 4229, 4229, 4229, 4229, 4229, 4229, 4229], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1284, "uuid": "69e19b38-b0d1-5c4d-80ac-a8c0dec788b0", "validation": {"axes": {"B": 9, "MP": 67, "NP": 603, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.017649661749601364, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "69e19b38-b0d1-5c4d-80ac-a8c0dec788b0"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f305ba292e5f76651f909c61f5c8172d41172dea8345a35d163eaef68ee1b2d3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d420f8f2c79bfe4eb8ef95bc70f3fa36300236214f5b735677b8cc9fa17ec412.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c9dfb122689153a3f4ab314ae925ae1d323811cafb1c53a85219c5a7ff5bbc00.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4239, 4239, 4239, 4239, 4239, 4239, 4239, 4239], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1294, "uuid": "a63bfb58-0989-5c8e-aaab-86d0bfdb63aa", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003035107161849737, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a63bfb58-0989-5c8e-aaab-86d0bfdb63aa"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/333adad0402756bc7e4f572fdf347c414ac24cf3e8072512c20148db499e7dab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4239, 4239, 4239, 4239, 4239, 4239, 4239, 4239], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1294, "uuid": "62656fa1-c8fc-5518-83df-7a0c6f4e0484", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.015137908048927784, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "62656fa1-c8fc-5518-83df-7a0c6f4e0484"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0473b002d64cc747f9cfef6057409ad680aae2fd52e0676a011592d0715fde37.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fc13fa87f23f370fd6f4985f7c67709512d0345846b2300cc5ae6495c27f4230.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/40f74539d8c69714ac5719b96efaa46c36eb88af5cac65f01ee32e1b05739ac3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4240, 4240, 4240, 4240, 4240, 4240, 4240, 4240], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1295, "uuid": "c10a27f1-c58f-596d-b5d6-f3caf2351041", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0030318088829517365, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c10a27f1-c58f-596d-b5d6-f3caf2351041"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d5a2b70b266bae404b5ec4b878db6690e0c93e9211bc24f1f2b6d7af2bf0379b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4240, 4240, 4240, 4240, 4240, 4240, 4240, 4240], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1295, "uuid": "ebc84a9b-10cf-5d0a-a81b-83fb68875e59", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.015067075379192829, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ebc84a9b-10cf-5d0a-a81b-83fb68875e59"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/89af38285528499a1dafb4241ccee91043de68812146c755637fbb517cb6ef55.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2cd24431363936e649e8b10d5cde4213885e24a7d0ddf002a1c8943484c19b08.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/34a40ea05144052ef89c4d342c5af6c86bf1979dd1bc7b4a830a8e3cebce4a93.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4241, 4241, 4241, 4241, 4241, 4241, 4241, 4241], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1296, "uuid": "d8d749cc-645a-5895-b587-f6329d60bc74", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.003018244169652462, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d8d749cc-645a-5895-b587-f6329d60bc74"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/bbda229ef448b2b275d1d2502c7373a4ea07540b3a2a1251abcf8edb9ae3b267.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4241, 4241, 4241, 4241, 4241, 4241, 4241, 4241], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1296, "uuid": "aaf9556a-8706-5895-8991-e85bc4ac1399", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.015205129981040955, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "aaf9556a-8706-5895-8991-e85bc4ac1399"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/67dcbc8069517ba33269421a676ac42cb66d3e99da1f49c0988c6f90a2fb1439.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9f651be2f9e2233023d893a16943bb492cfb2fef846774e23170f5a8ddc2aa98.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e96b725c38e2a2891910a1ac5121de03beb5b702f460a9c77578ba3c488f0932.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4242, 4242, 4242, 4242, 4242, 4242, 4242, 4242], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1297, "uuid": "93693b8f-fb78-57a9-aa52-61ee29861003", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0029192501679062843, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "93693b8f-fb78-57a9-aa52-61ee29861003"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8d9f703fff92cbe3299fe7758f283885809f8196b5012a605547b3d2f9081462.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/2e46e84e7c28414d3d392ea2f2b3a12ba0a27351eed2fa370140114a0cb14798.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4242, 4242, 4242, 4242, 4242, 4242, 4242, 4242], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1297, "uuid": "6854fc62-28ce-54df-b2b2-e98c2c055be3", "validation": {"axes": {"B": 8, "MP": 67, "NP": 536, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.015133010223507881, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6854fc62-28ce-54df-b2b2-e98c2c055be3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7e0901bc4fe6d6237b3643c3ac7a8bb613772c0c842afbe3c446989e4fe8763e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/22ee8e81eb826e6256a48bddbff703ad58dd84ad00f354bc529256cd64d71ef7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5bb75cc8ce216f468abd837a169a74dff62c7b2bf332f271e3b43840ea26176b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4251, 4251, 4251, 4251, 4251, 4251, 4251], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1306, "uuid": "efc8accc-c49e-58f6-bacb-318ec1612e27", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002805102150887251, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "efc8accc-c49e-58f6-bacb-318ec1612e27"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/a7d22eb242295bb73708489a42196989043f26bb5b120069a3e282840ae693ca.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4251, 4251, 4251, 4251, 4251, 4251, 4251], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1306, "uuid": "a72ed4c9-4673-59ca-a2d4-c408efb9c466", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.01639472460374236, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a72ed4c9-4673-59ca-a2d4-c408efb9c466"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9dd0df8257b8db4e9b308b6f743039231a35a4b31f72087e1bbf85b22994243d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a6c47803109f4d959df8af5e9920ffad91f576dc387c69cd32802cb6f32d2537.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6fd462fae2be9a04031f2da2ccd820a5d58c9f66a032e65efd667681d3413b93.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4252, 4252, 4252, 4252, 4252, 4252, 4252], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1307, "uuid": "f177bdb2-6029-5cd2-a9ba-901acef51319", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0028023207560181618, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "f177bdb2-6029-5cd2-a9ba-901acef51319"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/10975efbacdf52fa6817d71296d3c5cd95a9ad30d0045a854a5dee5eeb6f4226.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4252, 4252, 4252, 4252, 4252, 4252, 4252], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1307, "uuid": "19ba666b-20e1-5b37-978b-f9f72fb04cf8", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.013449644204229116, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "19ba666b-20e1-5b37-978b-f9f72fb04cf8"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6dd1cbae0c12ce9bd0ef1a71ffe0b9afb24649204350b7e9e650f9ad5479d926.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/117e522da5c183cfe8ae2c955ff9e2563b1e491e0489773f254469130fd735d7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/563a7d0457ea3b6aefb955681e68d68a1cc5f4b17629b3ef700426af2f9bba5d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4253, 4253, 4253, 4253, 4253, 4253, 4253], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1308, "uuid": "e06a85d0-b817-5df3-bbdc-cf2f435c2f36", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002769386861473322, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e06a85d0-b817-5df3-bbdc-cf2f435c2f36"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d4aa394f2554ade2099746ef4a5025c481a1d05c0519f70dbc5d5e96fd6db169.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4253, 4253, 4253, 4253, 4253, 4253, 4253], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1308, "uuid": "6d086432-475f-5475-87ab-97a1ebb1dc9a", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.013463621027767658, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6d086432-475f-5475-87ab-97a1ebb1dc9a"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/53db05b3a94d0c805fb620d15e4161b95a9ec3f11f2c231c7fe7ee0aa61965fd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2954c0b2f30f69e0b8d35489040a94e980ef306b9179a3466df04e8d7c8c7e67.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6a82902cd88c856a65dcc9ad3596a38cfac62b5986f7d5c7e6f8be2d20315448.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4254, 4254, 4254, 4254, 4254, 4254, 4254], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1309, "uuid": "c44d5e5c-aafa-5102-b3ef-b4d5a766f930", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0026933341287076473, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c44d5e5c-aafa-5102-b3ef-b4d5a766f930"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fd1f3af2b91fd8a36cf673111faa9b93ce0c7234c8de3c469c2304d34b7c7673.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5977fe4e0ffb080359a8f2d13e98b11c23099b40a1b284531d7bde1bda9b32ab.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4254, 4254, 4254, 4254, 4254, 4254, 4254], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1309, "uuid": "cd613028-cda4-5d7c-908f-1dcdbdb95ff9", "validation": {"axes": {"B": 7, "MP": 67, "NP": 469, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.01367177814245224, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cd613028-cda4-5d7c-908f-1dcdbdb95ff9"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3cd66aeccb44d9c7d6a5ad04c77259060ea74fb72bbf1b3901ddde0a11becbf1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/68aef16e47604790dc68f7b31960a705c37d334020a7aa8d1f9083b85c0abe0f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/a0186ba9599a9c75507a169e9b602d1fe6468ae9ac878de8e01fcf0837fa6aa7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4267, 4267, 4267, 4267, 4267, 4267], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1322, "uuid": "c0663e48-0132-5dd7-b443-862c34c2e945", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002536600921303034, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c0663e48-0132-5dd7-b443-862c34c2e945"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/400cc619857cc5928767811487cc65dea7f696a642bbc64f7142b19941ff1690.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4267, 4267, 4267, 4267, 4267, 4267], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1322, "uuid": "feff4d0d-8000-5f52-9b00-1de6ce58fe5d", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.011956102680414915, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "feff4d0d-8000-5f52-9b00-1de6ce58fe5d"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bfa46c5ba19a780ca9bca41f1506fdc7503585bd6505454e9f6dd7badf5fa169.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/599b15779b3abc6709fee78062e104bc5830a041f51f96ef6a1a7bddf93bb494.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8769d70e3c9dc817efd1376c960a58d7981df4ae847af8c29dcc068ff36f1ec3.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4268, 4268, 4268, 4268, 4268, 4268], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1323, "uuid": "6f718fe0-c788-572e-93ff-3ea7acfd8fa5", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0025196531787514687, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6f718fe0-c788-572e-93ff-3ea7acfd8fa5"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4077d17d16732449cc8a3f6cf2dded99dc981e053f340bc4ce0ec4c079d0c6f6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4268, 4268, 4268, 4268, 4268, 4268], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1323, "uuid": "bd5fcb96-2f7a-57c0-b0e7-75f722f192ca", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.01186472736299038, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bd5fcb96-2f7a-57c0-b0e7-75f722f192ca"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f091c38c6b6d08c9f08427a04376a5f3714d419397e1db67029c9212ed4a461c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f57d8c4dd15e9dfc69ee50e379e3f05b0c38af228d3b8939a618a42be5059730.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/61e3a8c40fc393068173c230bea53e7f32350ad02774e77ca0f704d79a016599.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4269, 4269, 4269, 4269, 4269, 4269], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1324, "uuid": "d53a75e2-4027-5b2f-bc1a-f7d6492ee0fc", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0024961568415164948, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d53a75e2-4027-5b2f-bc1a-f7d6492ee0fc"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/4c5528a4b7a88b5a3041c306b5fad8639563582f47b5b2e7c3610dab180c8cb1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4269, 4269, 4269, 4269, 4269, 4269], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1324, "uuid": "c6670bf1-9219-5cef-be39-7b8970e3ca5e", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.011734514031559229, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c6670bf1-9219-5cef-be39-7b8970e3ca5e"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/aba76e058aed3940a62f80821c392fffcf1976cd822ed2357e323da996996eb1.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/dece8d7a11b2a9a87ccbab55ad46615461d0917095d3e716c27c71b669bd445f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/f61842d659bddfc47d1cff7b25ffbb29361e6a8b284f13b7c564f8c4fb4b348a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4270, 4270, 4270, 4270, 4270, 4270], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1325, "uuid": "98ef5080-13b2-5711-b92c-c892d9264d45", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0025603012181818485, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "98ef5080-13b2-5711-b92c-c892d9264d45"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/76d19b60ae931686b1e506325fa87ae27795674eca55eea2982e6930374a774f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/fa4a878cb63640959c1a09926b07da8a826168f49582c7c4bae66a34f000fbbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4270, 4270, 4270, 4270, 4270, 4270], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1325, "uuid": "6a117b14-1211-52a5-83dd-acc1d69ef049", "validation": {"axes": {"B": 6, "MP": 67, "NP": 402, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.011678155977278948, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6a117b14-1211-52a5-83dd-acc1d69ef049"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3d7af82971b347755130c0dfbfdb98bf0644467fec472b29bccd7e70366acf5f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/79ccc10128de3df48a2884fb1b7f2bde681b6465fbd583aa0add6aafdeb4815a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c84db9884463f7f8b9ac6c21ed7b766e8388c74a28f0af54d864c75a587db172.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ed553459fc2abd106da36f5a939af71027739434825705c0eaa8fed06a80617b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4288, 4288, 4288, 4288, 4288], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1343, "uuid": "cb7d4f57-11d1-532a-8dd1-51f56fcbd0fd", "validation": {"axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002277795225381851, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cb7d4f57-11d1-532a-8dd1-51f56fcbd0fd"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 67, "NP": 335, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/8c2fac44e9438db974acd15f8a562f2dcba6cc1018259826a03adf8dbbb1b093.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/c84db9884463f7f8b9ac6c21ed7b766e8388c74a28f0af54d864c75a587db172.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4288, 4288, 4288, 4288, 4288], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1343, "uuid": "a60f36fd-d498-5eab-b118-536ac4ef945f", "validation": {"axes": {"B": 5, "MP": 67, "NP": 335, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.009924573823809624, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a60f36fd-d498-5eab-b118-536ac4ef945f"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/48bd7661cc43f3bfabe96c1f7b9c484a4a454a4c261dc555d06611a9c461b040.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/65c364f68472558c79644dbdf702869607f9b7e73880686d4486b45332b884fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c84db9884463f7f8b9ac6c21ed7b766e8388c74a28f0af54d864c75a587db172.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/93f4770ad1bc1e0403a5cb8e906d78b024f537b6b15109ded883aab0d4cad101.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4289, 4289, 4289, 4289, 4289], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1344, "uuid": "446b4fde-b5cd-500d-b183-2087fdf85353", "validation": {"axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002352232113480568, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "446b4fde-b5cd-500d-b183-2087fdf85353"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 68, "NP": 340, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b4628febb31f259c7b4bc90cfac8bc446ac888a0b7976530a1eb70667e1dad1f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d82a27196c3741e36772e68445de602d67f12e0f3885a300f49b1c5a45ed6418.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4289, 4289, 4289, 4289, 4289], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1344, "uuid": "e8c0cad2-4245-5c56-94cc-2ed45b58b76c", "validation": {"axes": {"B": 5, "MP": 68, "NP": 340, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010213776957243681, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e8c0cad2-4245-5c56-94cc-2ed45b58b76c"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/281eab2ec5ff66ec95144dc0fed733f1b9833699261bfc701b3883dca2fc00f0.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/026b24c848e9799d863998bbf60a0a1e5bc006cceb0420eb5da676468973ecbe.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c84db9884463f7f8b9ac6c21ed7b766e8388c74a28f0af54d864c75a587db172.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/646cf89248d054b275dddf23861b6eb7b63f8f0a72d04b8a701dceef6751fb05.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4290, 4290, 4290, 4290, 4290], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1345, "uuid": "05db939b-bd6b-5cf8-ad19-9f096fde1a67", "validation": {"axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0021924669854342937, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "05db939b-bd6b-5cf8-ad19-9f096fde1a67"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 68, "NP": 340, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/917d08163eaa2964311600bc46075c03413546931913417520f59674f79fb44f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d82a27196c3741e36772e68445de602d67f12e0f3885a300f49b1c5a45ed6418.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4290, 4290, 4290, 4290, 4290], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1345, "uuid": "1b1c535c-9395-5679-83e5-df31f077fdf4", "validation": {"axes": {"B": 5, "MP": 68, "NP": 340, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.010061345063149929, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1b1c535c-9395-5679-83e5-df31f077fdf4"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c5ce4cdc09ca674b79c4826451e3a1fb84f2b1fd9f5d2feb14fa7ea43f5ee65f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c3d2ed90574abd57e67518682d4a937c115fe9e7d784ac84dd7addaba96873ea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c84db9884463f7f8b9ac6c21ed7b766e8388c74a28f0af54d864c75a587db172.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/c560627c3d589236a2eed340d9de3cdf7756c66c088c8f252e65ee23425bc50b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4291, 4291, 4291, 4291, 4291], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1346, "uuid": "c50298a4-e300-5829-8dd1-b602cff57fa5", "validation": {"axes": {"B": 5, "MP": 67, "NP": 335, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0021637333557009697, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c50298a4-e300-5829-8dd1-b602cff57fa5"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 5, "MP": 68, "NP": 340, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/ab66adf64ac522695d43f930f75c0b252770f2118a163abe5f38550aada88584.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/d82a27196c3741e36772e68445de602d67f12e0f3885a300f49b1c5a45ed6418.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4291, 4291, 4291, 4291, 4291], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1346, "uuid": "cf368e98-831d-5daf-a8da-0ee994010600", "validation": {"axes": {"B": 5, "MP": 68, "NP": 340, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.010017660912126303, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "cf368e98-831d-5daf-a8da-0ee994010600"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cec487ec27c7f32985bb126eddb08a9967c964cc73b8890c9c898f53f1c34be4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b78fd199233bfe7cba6dad8f8208575da36c1e707d74adcc75f1afd822848a33.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d5c9b4952d81436916134e3b2815954beaf413536d877d51dce4986fc01eeffd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4319, 4319, 4319, 4319], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1374, "uuid": "fb8bea51-0c80-541c-9b9e-c6a217e60c4e", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002001275308430195, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fb8bea51-0c80-541c-9b9e-c6a217e60c4e"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/01c3ed198f844df62653f8a2b14758b79da7a29b68320605961c95a078880541.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4319, 4319, 4319, 4319], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1374, "uuid": "7cd6144a-b156-5be2-87e3-b182735328ab", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.008359548170119524, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7cd6144a-b156-5be2-87e3-b182735328ab"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/018c02bb2555d136f473adedca536a2a4826f503252bd76ee0ccee7e312c3542.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9dc6c5a64846ac26bbc24af35ad2958d5c0bc13e45fbb9f32ebe831637436de6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cff90c6c733a14735ba05ec0ff8f8edf418c7bb220671f855005be0982064a5d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4320, 4320, 4320, 4320], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1375, "uuid": "6e27b647-f5b4-5aeb-aba7-4be59491b047", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.002005794085562229, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "6e27b647-f5b4-5aeb-aba7-4be59491b047"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/56e0b51a6751a792258eda8185106e2d915f1b53e0441a0454e5088d87403834.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4320, 4320, 4320, 4320], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1375, "uuid": "926ad8d2-b65f-5ba8-8244-9d312916be50", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.008265772834420204, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "926ad8d2-b65f-5ba8-8244-9d312916be50"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/55a1f8b0e249fa5736d692293a0b5104614ad18e921407241729f126ee0420fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3cd97d2e25e44db2948fcb7c4e50fbe28cd3d5f1e83d3a3038b6669cfaf2091c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7b901ecd4f007a69acf708b8dc59cc5a7f4313da54342ff43c28d442264d3eb6.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4321, 4321, 4321, 4321], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1376, "uuid": "b321f630-e087-5c6e-98a5-050bcb45cfcd", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0019917436875402927, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b321f630-e087-5c6e-98a5-050bcb45cfcd"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/dda55fe1346a00aec132b4f41b45e8f7ed5a9fffa07cc2792c860137ba89e32b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4321, 4321, 4321, 4321], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1376, "uuid": "c25e380b-4853-582b-838d-c9457f484239", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.008297743275761604, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "c25e380b-4853-582b-838d-c9457f484239"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/98e7f6a792235650ee1a849d73b4370c5fa1b8330960b11c7a25a3eeba91e23e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3c913f34393030accb57b5254eebcb1251101e7b96380b2cb6bd644e5c0c818f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/cad8dc4712753b61ec50a0442cd6c3ab059db63db8dea2f3a9d0549760251d00.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4322, 4322, 4322, 4322], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1377, "uuid": "7baf6bff-c6be-5a6f-9772-7514190f1a93", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0019157081842422485, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "7baf6bff-c6be-5a6f-9772-7514190f1a93"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/de72f8c0398a0d530fe6e2384cba32543cb2ac8e1177e1aff910fb7f829bec97.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/578d31e6673372b5c68bd5399e20945d58ac9860df6bd67a1229ad3302ff7810.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4322, 4322, 4322, 4322], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1377, "uuid": "d39f703c-d13e-51d3-8444-4c41b8197900", "validation": {"axes": {"B": 4, "MP": 68, "NP": 272, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.00831991620361805, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d39f703c-d13e-51d3-8444-4c41b8197900"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b9a1aa5a0ec36bde112dadf3ac5b82dafca16ce7aa5a18e627fc23f0830d56b4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/350fda4c70096470ba8d73beab13db09479d114dac3216f014a7e4e14a8ee4bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/eb6217ab55d1b9222f2f15fdc4797105b78be6c9cbf25f127f62d9480437310d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4374, 4374, 4374], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1429, "uuid": "86565a09-8424-520c-a161-6209e6ce2795", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0017506349831819534, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "86565a09-8424-520c-a161-6209e6ce2795"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/9ca0f9a3202ad4b7e952569b4f7c6b9a98b84bcb3e2b4333c4daca3251713aed.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4374, 4374, 4374], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1429, "uuid": "b95a8954-3c98-576f-97c1-c82062f108ac", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.0076340921223163605, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b95a8954-3c98-576f-97c1-c82062f108ac"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fa67fa7614089d1bedcb878fd152e4ead90f973d7fdd16e5a765fdb51e860042.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/caa92927107cd956cbf65682322e70ea8339e8efd5ef8d7411f903f57c5cfcc8.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e624a93fd6787dab3ec21784ec0c8728498565573416b784721b786170fb37f9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4375, 4375, 4375], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1430, "uuid": "9fd9841d-86d6-5195-ab26-b47079f3af50", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016814540140330791, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9fd9841d-86d6-5195-ab26-b47079f3af50"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/cb8356cd3e4200cd69e19dcd75b670910ad5e090f16dfd13f5a8b2654a0cd6c4.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4375, 4375, 4375], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1430, "uuid": "a60407fc-f26e-5166-9165-d2784f0a8441", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.006682517938315868, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "a60407fc-f26e-5166-9165-d2784f0a8441"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/ddd50930792ccb3735964142975c2cb209158dce011d3e81c5aafe5cb9b1b07d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/52f8fdfb6637acb93aef64537e4efdfd2379202df2a82dfffd8e4fe477d76a65.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3dc2e99e58a7c95674d7d648d91867c71ed1f347ba02f1a4203c88bf1c6e4791.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4376, 4376, 4376], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1431, "uuid": "ac4881f5-e8c7-5dfd-8635-716978407fdc", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001703397836536169, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ac4881f5-e8c7-5dfd-8635-716978407fdc"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/515c863f21fd45eac6ce1088fb8ae3431fcd899403eced5be078102d417d67ea.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4376, 4376, 4376], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1431, "uuid": "1ca40483-b42b-5f92-b62c-202ae0d488c5", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.006619902327656746, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1ca40483-b42b-5f92-b62c-202ae0d488c5"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/be0be45172c799c8f53adcb5a25443fdb057f2c7103ea7f718884a06a62f12ee.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/067d686e1c900144cade5b9b1b9009d8e98e0b9097fc8a5fbc8cb3ce50af2379.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4ed6ac6e0c9d5b575a78c44b3cee62e8e5f42a68edf82bfeac97a4accb270922.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4377, 4377, 4377], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1432, "uuid": "223ce5d8-5634-5043-88ea-5f59aa71ba2b", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0016905390657484531, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "223ce5d8-5634-5043-88ea-5f59aa71ba2b"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/36c830c89940d11b814789ca19379a999f186ced4679f5161edca38e381dbf69.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/1dae96cf229c8c6377ed6e977365ae4857f5dd65a4815425ec8c6ea0ace9e4fc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4377, 4377, 4377], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1432, "uuid": "dab7ce78-081d-5344-9e42-184bf52ddcd2", "validation": {"axes": {"B": 3, "MP": 69, "NP": 207, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.006745555903762579, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "dab7ce78-081d-5344-9e42-184bf52ddcd2"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b0ca3765a288af2df98a465b2589323b105c6338e648cd0754a2eda4d821a45d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3f983a43a95a547848dc9ca3a1ad0c1eaa13eff3111e018ddf68c002f506b492.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4737ae53c7edd7c6ad67d6a073613ef6512dcd5ffc7bf9036c1fc66e2c6de6b2.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4382, 4382], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1437, "uuid": "1cabe9e2-7a83-5a46-95cf-916bef894fb2", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013946187682449818, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "1cabe9e2-7a83-5a46-95cf-916bef894fb2"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7296e6d3b731f208a2d92a14aa077b57c80e9c9649900b9ec2494f89555cbc09.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4382, 4382], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1437, "uuid": "ae0d4424-0e9d-50db-874b-18536b4a6e28", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.00480226892977953, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ae0d4424-0e9d-50db-874b-18536b4a6e28"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7bc01627386a7506227894182053041830d6cf04efc7b6d923068c301a47850f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/20aaa54004ad60a7acc9ae83dadb6ab4f3671fc3c624cf2a56c74e4a97abaa0c.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/6d469a26aedc2e5d312a50095e6bb195f408ffc5c616bd980623b7f18ddaec0e.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4383, 4383], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1438, "uuid": "b09f0987-4e0c-50de-b046-15ac9e6c10dc", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001390846911817789, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "b09f0987-4e0c-50de-b046-15ac9e6c10dc"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0fabd80b5f833bbec6c7d061575a913077e07750b2cdf21891ef956c38d39aaa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4383, 4383], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1438, "uuid": "fe75f570-2be4-519b-a73f-0059d0066497", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.004998452961444855, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "fe75f570-2be4-519b-a73f-0059d0066497"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/bbf2907e9c996e7ea3614a0508a5d2d1fe5c5fc147865bc2022c95598be0d487.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e4e5588a442801882f99dfa7ba1580560e1249b41698cca567f766af66a8301f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/9136e0cce8e1a1d9190f14d367cc65ff65729e5a3cc2ffd55afc32d3fe291b5f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4384, 4384], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1439, "uuid": "da23d286-b911-59c0-acff-87234878e081", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013775099068880081, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "da23d286-b911-59c0-acff-87234878e081"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/7aad11ed653c8f46dc0aa120e64fd1f241d2b513811a582acc45d7cd669d8b60.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4384, 4384], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1439, "uuid": "375edb3c-3113-5d12-974e-922e4bab6be3", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004880154971033335, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "375edb3c-3113-5d12-974e-922e4bab6be3"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/7b8461b4edbe6f21270d3eeeb08b76c40a77504a53b33984dc3fdc9b8322f6db.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/fc48f4c272644e3b0fe59f562abec3b809925d41163ca877c062181fcbc1c0ba.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e650cdf2b8aa72cdc8b35397a2cd3a8cc28d1ea9fa31d2a9dfad56304733f41f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4385, 4385], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1440, "uuid": "9c037788-b2b6-5de7-a67d-3900df0c8d6f", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0013828580267727375, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "9c037788-b2b6-5de7-a67d-3900df0c8d6f"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e2d5ea210289c386ee9c43547ca07cbd8242d04ac4acba266613acf1433b0c92.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/5f6ad3f9eb2aea323f9ccac7e5049b4bd7721d5fe7363ba63125498aa3b9b7bd.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4385, 4385], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1440, "uuid": "e560f34c-cf2d-5d7a-bb54-49e89060ee3b", "validation": {"axes": {"B": 2, "MP": 69, "NP": 138, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.004971201065927744, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "e560f34c-cf2d-5d7a-bb54-49e89060ee3b"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8582159d83612d89652af7f3be774888dbdb5163a1cead3cab3386f59e11a863.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/89a78cfd8a29df94000fc57b86d0f4c9e4a7edd939bd30772752c1d85428c0ac.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/4713557a08a6eaedecbccc77e94766e6dfd30e65448d4172f66a0334f76e23e9.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4507], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1562, "uuid": "d79fa99d-c281-5ad4-9d1d-8a1038168727", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001096110325306654, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d79fa99d-c281-5ad4-9d1d-8a1038168727"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/44fdac32a724be75fcd9ab21f667e3a29518da319fec8717921e0732b74d4465.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4507], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1562, "uuid": "4b575c74-f88c-5838-ac91-223a831eb5d4", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.00350704463198781, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "4b575c74-f88c-5838-ac91-223a831eb5d4"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/2d4c349e0dc54585b94b108790179d8ffd986ae518a9d89fca9e47abb907faca.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/d130edb2520645a35e86300b05311df1415c11ea4b4872accd54fa7b3b82f789.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/8151a1ae6d5e4cbd02c75bd6034aafbe300532b760e20dde171b801e98330f1f.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4508], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1563, "uuid": "beb95180-1268-5e3b-8185-f9c4eda90487", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.0010786708444356918, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "beb95180-1268-5e3b-8185-f9c4eda90487"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/aa646c7bf831d8846466ecf213f7864796bfb448348123e40a85d5adcd9acff7.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4508], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1563, "uuid": "bef70f95-aebc-52d6-b88e-adbfc11648dc", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003161379136145115, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "bef70f95-aebc-52d6-b88e-adbfc11648dc"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/1764930972cd3752bf216393cf9e2fb284778339d254aa495458d7817323e5ec.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/3dc2692d0757bc6dc7ea2b7d2059e43e1b5e5ecfac5f22686650e2701b48f420.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/b05f1165d30dbb8692af4de65c879f78258d8cff5197463f907c8eba1c21d928.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4509], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1564, "uuid": "d3c8ace0-1356-5ec7-b48f-14659572c8b7", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001049961894750595, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "d3c8ace0-1356-5ec7-b48f-14659572c8b7"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/0b54bb4223474e3259acd641efd739edd755e3f50e659373299bba4e28c4d6aa.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4509], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1564, "uuid": "41ad8b2d-d282-53c6-b035-c92217b58ad9", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.015625], "seconds": 0.0031108888797461987, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "41ad8b2d-d282-53c6-b035-c92217b58ad9"}}
{"adapter": "Retain all live pages byte-for-byte and remap their page IDs; omit unreachable allocator pages.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 16, 1, 128], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "engine": "SGLang", "engine_block_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/0a0c37812d3c5236ed209a6ec0c88132fdf1437c739c702494f3cab51b737a2d.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/664e7ff4fcacf6e9ebca3b8596fb0364f795e94687c461f2294d073a9654912a.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QSAIndexer.select_decode_tokens", "original_physical_page_ids": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "output_adapter": "Serialize the engine's selected set in increasing logical-ID order; preserve membership and -1 suffix padding.", "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "padded_query": {"path": "blob/workloads/qsa_index_topk_decode_qwen38_h4_d128_topk512_c4/94963b985da74f773b2f566f737b6599a4a78e7314cb1d584fccd92bf0574efc.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "query_adapter": "Remove verified zero query heads inserted for the backend's eight-head tile.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4510], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1565, "uuid": "5d1b177b-fffe-5d60-915f-f25a8a14baff", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 16}, "definition": "qsa_index_topk_decode_qwen38_h4_d128_topk512_c4", "device": "cuda", "max_absolute_errors": [0], "seconds": 0.001022164709866047, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "5d1b177b-fffe-5d60-915f-f25a8a14baff"}}
{"adapter": "Snapshot after the current-token KV write; retain every live page and remap physical IDs.", "agentx_dataset": "semianalysisai/cc-traces-weka-062126", "agentx_revision": "23f152f6f0f9399a85901b89a6458def0ef16729", "allocated_cache_shape": [83743, 64, 2, 256], "axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "engine": "SGLang", "engine_version": "0.0.0.dev1+g684fea8af", "expected_output": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/b0fe872c2893ab11d7af19a895c8df2f1f5f09b7bcf4eca6cab7b267e7323490.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "fixture_sha256": "910fedaa22efdfa0ef4df0b2f2d13385622db054c30f1cc18438b300a7f9aba4", "layer_id": 3, "members": [{"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1906, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6521.707, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[2]", "model": "claude-opus-4-8", "output_tokens": 900, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6522.024, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1375, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.483, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_005_c9c37a94", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[20].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1430, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6523.487, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1185, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.526, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1295, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6525.97, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.284, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_006_534be3c4", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[21].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1307, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6526.319, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1250, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.072, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1323, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.583, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1282, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6528.979, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_007_69c5d16b", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[22].requests[4]", "model": "claude-opus-4-8", "output_tokens": 1244, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.227, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1563, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.367, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1438, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6529.86, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_008_8261a9b9", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[23].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1177, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6530.028, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 33024, "locator": "requests[25].requests[0]", "model": "claude-opus-4-8", "output_tokens": 352, "reusable_prefix_tokens": 29760, "reusable_prefix_tokens_256k": 29760, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.358, "uncached_tokens": 3264, "uncached_tokens_256k": 3264}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[1]", "model": "claude-opus-4-8", "output_tokens": 755, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6534.599, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1056, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.03, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[3]", "model": "claude-opus-4-8", "output_tokens": 936, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.145, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_009_46d5905c", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[24].requests[4]", "model": "claude-opus-4-8", "output_tokens": 955, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6535.159, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[1]", "model": "claude-opus-4-8", "output_tokens": 1069, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6537.877, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[2]", "model": "claude-opus-4-8", "output_tokens": 1344, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.263, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}, {"block_size": 64, "branch_id": "subagent_010_bb8f8581", "hash_id_scope": "local", "in_256k_variant": true, "input_tokens": 2944, "locator": "requests[25].requests[3]", "model": "claude-opus-4-8", "output_tokens": 1058, "reusable_prefix_tokens": 0, "reusable_prefix_tokens_256k": 0, "role": "subagent", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586, "uncached_tokens": 2944, "uncached_tokens_256k": 2944}], "model": "nvidia/Qwen3.8-Flash-Next-NVFP4", "model_revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47", "operation": "QwenSparseAttnBackend.forward_decode", "original_physical_page_ids": {"path": "blob/workloads/qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4/e705e742695c768f54ad864d2c6578032430e2404a0931a0d8dece027b5bbd5b.safetensors", "tensor_key": "tensor", "type": "safetensors"}, "overlap_profile": {"batch_size": 23, "label": "overlap_p100", "model": "claude-opus-4-8", "session_id": "5cbd632dae8626a5743d4a3ce38bd74ff9dc", "t": 6538.586}, "prompt_reconstruction": "Public AgentX hash blocks select deterministic spans of an authored code corpus; original prompt text is unavailable. Input IDs preserve exact recorded block lengths.", "request_id": "qwen_next_overlap_p100", "scheduler_mode": "decode", "scheduler_sequence_lengths": [4510], "scheduling": "Submit the observed overlapping requests together; record the actual engine batch, without claiming to reproduce their original timing.", "server_command": ["python", "-m", "sglang.launch_server", "--model-path", "/task/models/qwen-next-nvfp4", "--host", "127.0.0.1", "--port", "30200", "--dtype", "bfloat16", "--moe-runner-backend", "flashinfer_cutlass", "--ple-offload-embedding", "--linear-attn-decode-backend", "triton", "--linear-attn-prefill-backend", "triton", "--mamba-ssm-dtype", "float32", "--disable-cuda-graph", "--disable-overlap-schedule", "--chunked-prefill-size", "4096", "--mem-fraction-static", "0.8", "--max-running-requests", "24", "--max-mamba-cache-size", "120"], "serving_environment": {"SGLANG_BUILD_COMMIT": "684fea8af3b11a5b51ec9966c02959a32ad3421c", "SGLANG_BUILD_URL": "https://github.com/sgl-project/sglang/actions/runs/38489800016", "SGLANG_CACHE_DIR": "/task/cache/sglang", "SGLANG_IMAGE_TAG": "lmsysorg/sglang:nightly-dev-20261014-684fea8a", "SGLANG_RUST_BUILD_MODE": "never"}, "step": 1565, "uuid": "ecd71e48-73ea-5a55-bcba-553ea6141d4c", "validation": {"axes": {"B": 1, "MP": 71, "NP": 71, "P": 64}, "definition": "qsa_sparse_gqa_decode_qwen38_h24_kv2_d256_topk512_c4", "device": "cuda", "max_absolute_errors": [0.0078125], "seconds": 0.003144789021462202, "status": "passed", "tolerances": {"bfloat16": {"atol": 0.005, "rtol": 0.016}, "float32": {"atol": 2e-05, "rtol": 0.0001}, "gdn_prefill_state": {"max_head_relative_rms": 0.02, "reason": "The engine's chunk algorithm rounds intermediate matrix operands to BF16 although its final state is stored in FP32.", "relative_max": 0.015, "relative_rms": 0.01}, "integer": "exact"}, "torch": "2.14.1+cu130", "uuid": "ecd71e48-73ea-5a55-bcba-553ea6141d4c"}}
{"uuid": "901e5104-dccb-4c3f-ae13-ef4d31a4d456", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "901e5104-dccb-4c3f-ae13-ef4d31a4d456", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "901e5104-dccb-4c3f-ae13-ef4d31a4d456", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 16.271268669515848, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a5714b69-525c-4b95-bb7a-a0f9770c2f48", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a5714b69-525c-4b95-bb7a-a0f9770c2f48", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a5714b69-525c-4b95-bb7a-a0f9770c2f48", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.003432820551097393, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "3daa0974-293c-4414-b3c2-1f04368c1189", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "3daa0974-293c-4414-b3c2-1f04368c1189", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "3daa0974-293c-4414-b3c2-1f04368c1189", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0018753791227936745, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "22d5cef5-4f30-4f43-9d5f-0e9e95dc2201", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "22d5cef5-4f30-4f43-9d5f-0e9e95dc2201", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "22d5cef5-4f30-4f43-9d5f-0e9e95dc2201", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.002298467792570591, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "aed4bdd4-3139-4a1b-ae2f-aab8d4ba4090", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "aed4bdd4-3139-4a1b-ae2f-aab8d4ba4090", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "aed4bdd4-3139-4a1b-ae2f-aab8d4ba4090", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0017917631193995476, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "49a125e5-edf0-492e-a8b3-3676d14adaa3", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "49a125e5-edf0-492e-a8b3-3676d14adaa3", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "49a125e5-edf0-492e-a8b3-3676d14adaa3", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0017240457236766815, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "798635cf-d424-4343-a959-c96b0c0e81fb", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "798635cf-d424-4343-a959-c96b0c0e81fb", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "798635cf-d424-4343-a959-c96b0c0e81fb", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.001755768433213234, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "5716e24a-3f55-411a-bcd8-e6b677b1ca7e", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "5716e24a-3f55-411a-bcd8-e6b677b1ca7e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "5716e24a-3f55-411a-bcd8-e6b677b1ca7e", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0016884058713912964, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "4c7df22f-70ef-4494-864f-6f10209ab0f3", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "4c7df22f-70ef-4494-864f-6f10209ab0f3", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "4c7df22f-70ef-4494-864f-6f10209ab0f3", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0016648145392537117, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d0e91dea-aa1b-46c8-a67c-b2814f5a1725", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d0e91dea-aa1b-46c8-a67c-b2814f5a1725", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d0e91dea-aa1b-46c8-a67c-b2814f5a1725", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 1}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0016696881502866745, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ec9d2340-6d13-40e4-a6fe-4483a1cacd0d", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ec9d2340-6d13-40e4-a6fe-4483a1cacd0d", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ec9d2340-6d13-40e4-a6fe-4483a1cacd0d", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.003921466879546642, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "9f238670-9a56-4ab9-94f9-555755f32205", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "9f238670-9a56-4ab9-94f9-555755f32205", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "9f238670-9a56-4ab9-94f9-555755f32205", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.003905275836586952, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "6d929bff-d051-4e1d-acbf-7d0cc13f6dc8", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "6d929bff-d051-4e1d-acbf-7d0cc13f6dc8", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "6d929bff-d051-4e1d-acbf-7d0cc13f6dc8", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0038549499586224556, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "494844dc-80e9-41e9-9fe2-3d2618fdef64", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "494844dc-80e9-41e9-9fe2-3d2618fdef64", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "494844dc-80e9-41e9-9fe2-3d2618fdef64", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0038531990721821785, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "26f760ab-c286-41bf-8c37-3bc5df4c98fc", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "26f760ab-c286-41bf-8c37-3bc5df4c98fc", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "26f760ab-c286-41bf-8c37-3bc5df4c98fc", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0038051651790738106, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "deecf42e-dfae-4061-936f-0af4d892c231", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "deecf42e-dfae-4061-936f-0af4d892c231", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "deecf42e-dfae-4061-936f-0af4d892c231", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.004169170744717121, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "9a92acbc-6104-48c6-8176-601feee30001", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "9a92acbc-6104-48c6-8176-601feee30001", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "9a92acbc-6104-48c6-8176-601feee30001", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0038183266296982765, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "fa90b213-3ad2-4a42-9173-97fbf9e2e809", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "fa90b213-3ad2-4a42-9173-97fbf9e2e809", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "fa90b213-3ad2-4a42-9173-97fbf9e2e809", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0038338247686624527, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "2640f1e3-4b02-4041-bdc1-59a28e0b9954", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "2640f1e3-4b02-4041-bdc1-59a28e0b9954", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "2640f1e3-4b02-4041-bdc1-59a28e0b9954", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.007158014923334122, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "7464a1d6-b0dd-4061-b15a-ebdabf47e351", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "7464a1d6-b0dd-4061-b15a-ebdabf47e351", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "7464a1d6-b0dd-4061-b15a-ebdabf47e351", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.00710301473736763, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "25d1f606-3e9e-49e6-a3f0-991092c7a845", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "25d1f606-3e9e-49e6-a3f0-991092c7a845", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "25d1f606-3e9e-49e6-a3f0-991092c7a845", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.007128657773137093, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "fe2e6584-17d0-4025-ad44-db9e507f6bed", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "fe2e6584-17d0-4025-ad44-db9e507f6bed", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "fe2e6584-17d0-4025-ad44-db9e507f6bed", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.007044089958071709, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ca62112d-68e8-4eb7-a318-26c48d256e10", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ca62112d-68e8-4eb7-a318-26c48d256e10", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ca62112d-68e8-4eb7-a318-26c48d256e10", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.007242261432111263, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "61dfd334-971b-4a44-a8a8-f00950192745", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "61dfd334-971b-4a44-a8a8-f00950192745", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "61dfd334-971b-4a44-a8a8-f00950192745", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0074678147211670876, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "abec5d32-2409-4412-8683-472f3a091a54", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "abec5d32-2409-4412-8683-472f3a091a54", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "abec5d32-2409-4412-8683-472f3a091a54", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 8}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0071442630141973495, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "76eec66d-0ada-4c35-bf21-a07247ad7f05", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "76eec66d-0ada-4c35-bf21-a07247ad7f05", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "76eec66d-0ada-4c35-bf21-a07247ad7f05", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014222866855561733, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f0507025-1db0-48de-b3d6-4faad5632558", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f0507025-1db0-48de-b3d6-4faad5632558", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f0507025-1db0-48de-b3d6-4faad5632558", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014477045275270939, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d1508f6d-81b7-4caf-947b-b48f612a3061", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d1508f6d-81b7-4caf-947b-b48f612a3061", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d1508f6d-81b7-4caf-947b-b48f612a3061", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014274150133132935, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "63d4cb66-d6f6-4ceb-8083-0c0fb4abece9", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "63d4cb66-d6f6-4ceb-8083-0c0fb4abece9", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "63d4cb66-d6f6-4ceb-8083-0c0fb4abece9", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014356067404150963, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ef6bb01d-1294-4a06-97b2-604ec944b4c0", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ef6bb01d-1294-4a06-97b2-604ec944b4c0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ef6bb01d-1294-4a06-97b2-604ec944b4c0", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014112823642790318, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d0acf04a-5919-4bf3-95ca-94fbef5786c0", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d0acf04a-5919-4bf3-95ca-94fbef5786c0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d0acf04a-5919-4bf3-95ca-94fbef5786c0", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.01457969006150961, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ac83f1e4-982e-41a5-8c02-7ccc14c728d4", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ac83f1e4-982e-41a5-8c02-7ccc14c728d4", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ac83f1e4-982e-41a5-8c02-7ccc14c728d4", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014064210467040539, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a8c8beff-e414-4580-b2f7-e5b8f13bc269", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a8c8beff-e414-4580-b2f7-e5b8f13bc269", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a8c8beff-e414-4580-b2f7-e5b8f13bc269", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.028617889620363712, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "40940a4c-1348-416d-aea1-f3183548229b", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "40940a4c-1348-416d-aea1-f3183548229b", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "40940a4c-1348-416d-aea1-f3183548229b", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.028055261820554733, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "0915952b-5887-43aa-9807-a0980e7a78f6", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "0915952b-5887-43aa-9807-a0980e7a78f6", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "0915952b-5887-43aa-9807-a0980e7a78f6", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.02969113551080227, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d38dde79-79f9-4c2d-8bbe-05ee1be48583", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d38dde79-79f9-4c2d-8bbe-05ee1be48583", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d38dde79-79f9-4c2d-8bbe-05ee1be48583", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.028827985748648643, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "388f79b1-13e9-44e6-b12e-c8197989a924", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "388f79b1-13e9-44e6-b12e-c8197989a924", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "388f79b1-13e9-44e6-b12e-c8197989a924", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.028198459185659885, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "80288532-6967-4b8e-b5bc-c31ce0a208d5", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "80288532-6967-4b8e-b5bc-c31ce0a208d5", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "80288532-6967-4b8e-b5bc-c31ce0a208d5", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.02865928690880537, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "773a4157-01fa-4637-90be-8604e0b40526", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "773a4157-01fa-4637-90be-8604e0b40526", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "773a4157-01fa-4637-90be-8604e0b40526", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 32}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.028739776462316513, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "53385c7f-393d-41db-aec8-5b9eb5bf35d1", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "53385c7f-393d-41db-aec8-5b9eb5bf35d1", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "53385c7f-393d-41db-aec8-5b9eb5bf35d1", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.04420843441039324, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "45e7697a-bae5-42bd-9f56-959c10ff681f", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "45e7697a-bae5-42bd-9f56-959c10ff681f", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "45e7697a-bae5-42bd-9f56-959c10ff681f", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0434917937964201, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "34403e5b-8551-4d44-95b0-21532c4eb839", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "34403e5b-8551-4d44-95b0-21532c4eb839", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "34403e5b-8551-4d44-95b0-21532c4eb839", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.042750352062284946, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "b91cc4aa-048d-4e05-9c0c-8a153ec997f8", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "b91cc4aa-048d-4e05-9c0c-8a153ec997f8", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "b91cc4aa-048d-4e05-9c0c-8a153ec997f8", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.043537819758057594, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "4c38f0e4-aabf-4742-90b9-24231effd96f", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "4c38f0e4-aabf-4742-90b9-24231effd96f", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "4c38f0e4-aabf-4742-90b9-24231effd96f", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.043353582732379436, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "be91d78c-1114-4563-b0aa-e0682ccc09bb", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "be91d78c-1114-4563-b0aa-e0682ccc09bb", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "be91d78c-1114-4563-b0aa-e0682ccc09bb", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.04277325887233019, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "3803367a-f140-432a-9784-43fe5d5f64d0", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "3803367a-f140-432a-9784-43fe5d5f64d0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "3803367a-f140-432a-9784-43fe5d5f64d0", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 48}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.042964029125869274, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "eaf0a285-447c-4432-8e68-d287acc3cb08", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "eaf0a285-447c-4432-8e68-d287acc3cb08", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "eaf0a285-447c-4432-8e68-d287acc3cb08", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.058285756036639214, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f4fd2171-c869-440c-b199-f403e3c6b788", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f4fd2171-c869-440c-b199-f403e3c6b788", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f4fd2171-c869-440c-b199-f403e3c6b788", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.057347020134329796, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "b1931bc6-b8b6-44da-a83b-aa6cc4d4c275", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "b1931bc6-b8b6-44da-a83b-aa6cc4d4c275", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "b1931bc6-b8b6-44da-a83b-aa6cc4d4c275", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.057033597491681576, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "8d0c5789-550a-4216-a471-202b2655a4e8", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "8d0c5789-550a-4216-a471-202b2655a4e8", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "8d0c5789-550a-4216-a471-202b2655a4e8", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.057457820512354374, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "9eef9bc7-5972-437e-b488-853c6dacb470", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "9eef9bc7-5972-437e-b488-853c6dacb470", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "9eef9bc7-5972-437e-b488-853c6dacb470", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.05687881261110306, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "58571e49-29ee-4a53-bf1d-a7a2363e9db3", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "58571e49-29ee-4a53-bf1d-a7a2363e9db3", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "58571e49-29ee-4a53-bf1d-a7a2363e9db3", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.05689228139817715, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ef2cf980-6977-49d1-a6f8-7247becf0273", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ef2cf980-6977-49d1-a6f8-7247becf0273", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ef2cf980-6977-49d1-a6f8-7247becf0273", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.05679920595139265, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "42963acb-f2f5-4ada-9205-3931cd26fa44", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "42963acb-f2f5-4ada-9205-3931cd26fa44", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "42963acb-f2f5-4ada-9205-3931cd26fa44", "definition": "gdn_decode_qk4_v8_d128_k_last", "axes": {"batch_size": 64}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.055893903598189354, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "77daf91d-0660-4c4b-8c32-336a69281cd9", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 6, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "77daf91d-0660-4c4b-8c32-336a69281cd9", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "77daf91d-0660-4c4b-8c32-336a69281cd9", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 6, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.34122417494654655, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ba08a83e-e151-4e16-bc70-abee6851604c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 134, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ba08a83e-e151-4e16-bc70-abee6851604c", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ba08a83e-e151-4e16-bc70-abee6851604c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 134, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.02165132388472557, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "c7846f96-c9ff-44d3-94ab-5eab99a1431b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 134, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "c7846f96-c9ff-44d3-94ab-5eab99a1431b", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "c7846f96-c9ff-44d3-94ab-5eab99a1431b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 134, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.021006835624575615, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d0ce7b5d-49e2-4a0b-b2ce-a087139b7d6b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 32, "len_cu_seqlens": 33}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d0ce7b5d-49e2-4a0b-b2ce-a087139b7d6b", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d0ce7b5d-49e2-4a0b-b2ce-a087139b7d6b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 32, "len_cu_seqlens": 33}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2272192491218448, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "5b8a0e4b-5ed7-45d0-aec3-c2c154e515b1", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 34, "len_cu_seqlens": 35}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "5b8a0e4b-5ed7-45d0-aec3-c2c154e515b1", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "5b8a0e4b-5ed7-45d0-aec3-c2c154e515b1", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 34, "len_cu_seqlens": 35}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2189221857115626, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "5d3fc66a-6bba-4602-8319-a1da07045354", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 34, "len_cu_seqlens": 35}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "5d3fc66a-6bba-4602-8319-a1da07045354", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "5d3fc66a-6bba-4602-8319-a1da07045354", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 34, "len_cu_seqlens": 35}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.213029868900776, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "4b6143dd-0e5f-499f-93cb-076d9635bcd0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 4124, "num_seqs": 15, "len_cu_seqlens": 16}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "4b6143dd-0e5f-499f-93cb-076d9635bcd0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "4b6143dd-0e5f-499f-93cb-076d9635bcd0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 4124, "num_seqs": 15, "len_cu_seqlens": 16}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.6143626067787409, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "5835a2bc-8d60-43fc-b1ed-d4729ea62693", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 32, "len_cu_seqlens": 33}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "5835a2bc-8d60-43fc-b1ed-d4729ea62693", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "5835a2bc-8d60-43fc-b1ed-d4729ea62693", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 32, "len_cu_seqlens": 33}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2122697094455361, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "cc310f94-60ea-4dcd-9e63-135dea91b83a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 25, "len_cu_seqlens": 26}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "cc310f94-60ea-4dcd-9e63-135dea91b83a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "cc310f94-60ea-4dcd-9e63-135dea91b83a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 25, "len_cu_seqlens": 26}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.226369529031217, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d49df0b2-7838-4865-9a51-7ec7877fe27f", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 48, "len_cu_seqlens": 49}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d49df0b2-7838-4865-9a51-7ec7877fe27f", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d49df0b2-7838-4865-9a51-7ec7877fe27f", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 48, "len_cu_seqlens": 49}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2082636235281825, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "e9e1e445-b41a-404b-a20f-bc869acc6cf0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 38, "len_cu_seqlens": 39}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "e9e1e445-b41a-404b-a20f-bc869acc6cf0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "e9e1e445-b41a-404b-a20f-bc869acc6cf0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 38, "len_cu_seqlens": 39}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2139668138697743, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "b8c8dc3c-3dfe-4334-a784-c477213e2f1a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 38, "len_cu_seqlens": 39}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "b8c8dc3c-3dfe-4334-a784-c477213e2f1a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "b8c8dc3c-3dfe-4334-a784-c477213e2f1a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 38, "len_cu_seqlens": 39}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2213064078241587, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a9540651-862b-44af-ad75-57e2d9caf37a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 20, "len_cu_seqlens": 21}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a9540651-862b-44af-ad75-57e2d9caf37a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a9540651-862b-44af-ad75-57e2d9caf37a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 20, "len_cu_seqlens": 21}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2171437125653028, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "06f21bb1-6cbd-4d55-b620-fb4d62181a71", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 56, "len_cu_seqlens": 57}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "06f21bb1-6cbd-4d55-b620-fb4d62181a71", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "06f21bb1-6cbd-4d55-b620-fb4d62181a71", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 56, "len_cu_seqlens": 57}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2162983929738402, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "c2931c92-bbe9-43d9-bc55-9aed77901266", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 37, "len_cu_seqlens": 38}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "c2931c92-bbe9-43d9-bc55-9aed77901266", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "c2931c92-bbe9-43d9-bc55-9aed77901266", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 37, "len_cu_seqlens": 38}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2173201236873865, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "618df04a-540c-42fd-b2ce-285dfdf78422", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 35, "len_cu_seqlens": 36}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "618df04a-540c-42fd-b2ce-285dfdf78422", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "618df04a-540c-42fd-b2ce-285dfdf78422", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 35, "len_cu_seqlens": 36}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2180588580667973, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "26244fb4-d295-4d4b-bf61-235d17fd0149", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 43, "len_cu_seqlens": 44}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "26244fb4-d295-4d4b-bf61-235d17fd0149", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "26244fb4-d295-4d4b-bf61-235d17fd0149", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 43, "len_cu_seqlens": 44}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.221663354896009, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a2629e02-611f-4264-8ee7-5c431f5fee4f", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 43, "len_cu_seqlens": 44}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a2629e02-611f-4264-8ee7-5c431f5fee4f", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a2629e02-611f-4264-8ee7-5c431f5fee4f", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 43, "len_cu_seqlens": 44}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2279813839122653, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "9a5d694b-7d4c-4ee6-8315-a13053ab6f92", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 57, "len_cu_seqlens": 58}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "9a5d694b-7d4c-4ee6-8315-a13053ab6f92", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "9a5d694b-7d4c-4ee6-8315-a13053ab6f92", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 57, "len_cu_seqlens": 58}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2244162587448955, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "410794d4-6f70-4bb4-aed1-4669be9a610f", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 39, "len_cu_seqlens": 40}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "410794d4-6f70-4bb4-aed1-4669be9a610f", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "410794d4-6f70-4bb4-aed1-4669be9a610f", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 8192, "num_seqs": 39, "len_cu_seqlens": 40}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 1.2256116531789303, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "7ba9d519-7c0b-4eac-b2ad-ae403dce2f95", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 3999, "num_seqs": 13, "len_cu_seqlens": 14}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "7ba9d519-7c0b-4eac-b2ad-ae403dce2f95", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "7ba9d519-7c0b-4eac-b2ad-ae403dce2f95", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 3999, "num_seqs": 13, "len_cu_seqlens": 14}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.5933553474023938, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "043e74e4-0297-4988-808a-0090be48a90e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 32, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "043e74e4-0297-4988-808a-0090be48a90e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "043e74e4-0297-4988-808a-0090be48a90e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 32, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.005895373411476612, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ef9515b6-ad88-4a3e-bd89-31384ddd53ad", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 16, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ef9515b6-ad88-4a3e-bd89-31384ddd53ad", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ef9515b6-ad88-4a3e-bd89-31384ddd53ad", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 16, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0034635253250598907, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f622a11d-8b79-40b6-9949-5509afe827c8", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 16, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f622a11d-8b79-40b6-9949-5509afe827c8", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f622a11d-8b79-40b6-9949-5509afe827c8", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 16, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0033882055431604385, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "1cf8e175-9f87-4add-984a-6c87a380fe9d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 40, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "1cf8e175-9f87-4add-984a-6c87a380fe9d", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "1cf8e175-9f87-4add-984a-6c87a380fe9d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 40, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.007024860940873623, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "9343fd82-a06d-493a-9918-1044b0c1cbd1", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 49, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "9343fd82-a06d-493a-9918-1044b0c1cbd1", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "9343fd82-a06d-493a-9918-1044b0c1cbd1", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 49, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.008356481790542603, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f4926229-e4ba-49a3-8b2f-df2ec79c333b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1377, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f4926229-e4ba-49a3-8b2f-df2ec79c333b", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f4926229-e4ba-49a3-8b2f-df2ec79c333b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1377, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.206788363866508, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "109addb1-15e0-4ff2-9b39-df3e79746af0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 3271, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "109addb1-15e0-4ff2-9b39-df3e79746af0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "109addb1-15e0-4ff2-9b39-df3e79746af0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 3271, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.4837587783113122, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "c5257f65-c411-4dbb-9dc1-ad4abcf00254", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 983, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "c5257f65-c411-4dbb-9dc1-ad4abcf00254", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "c5257f65-c411-4dbb-9dc1-ad4abcf00254", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 983, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.14653201960027218, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f5619793-6267-4a96-8a23-6c6a937183e7", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 983, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f5619793-6267-4a96-8a23-6c6a937183e7", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f5619793-6267-4a96-8a23-6c6a937183e7", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 983, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.14694944210350513, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "fdf5f1f4-1135-4d3f-99e4-8965958317fc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 46, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "fdf5f1f4-1135-4d3f-99e4-8965958317fc", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "fdf5f1f4-1135-4d3f-99e4-8965958317fc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 46, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.00795680470764637, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "87bff084-1b7a-478e-99fd-5952c989d80e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 202, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "87bff084-1b7a-478e-99fd-5952c989d80e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "87bff084-1b7a-478e-99fd-5952c989d80e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 202, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.03042548894882202, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "e92dafeb-3d50-4bce-90c3-6d0172e189d3", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 239, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "e92dafeb-3d50-4bce-90c3-6d0172e189d3", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "e92dafeb-3d50-4bce-90c3-6d0172e189d3", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 239, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0360850915312767, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "1d0cc342-9e3d-41e3-9ade-532551f86c17", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 294, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "1d0cc342-9e3d-41e3-9ade-532551f86c17", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "1d0cc342-9e3d-41e3-9ade-532551f86c17", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 294, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.043932353146374226, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "1b441950-6f77-4565-bada-12294714fb26", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 91, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "1b441950-6f77-4565-bada-12294714fb26", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "1b441950-6f77-4565-bada-12294714fb26", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 91, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014433187432587147, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "19c6ab20-df55-4b2c-b9a7-d76870b50e3e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 91, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "19c6ab20-df55-4b2c-b9a7-d76870b50e3e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "19c6ab20-df55-4b2c-b9a7-d76870b50e3e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 91, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014619696885347366, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "25d9c14d-90ad-442d-8b0f-9452ad198832", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 132, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "25d9c14d-90ad-442d-8b0f-9452ad198832", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "25d9c14d-90ad-442d-8b0f-9452ad198832", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 132, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.020527238957583904, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "3215fe5f-4a3b-4eb6-af20-4e17368d87a9", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 341, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "3215fe5f-4a3b-4eb6-af20-4e17368d87a9", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "3215fe5f-4a3b-4eb6-af20-4e17368d87a9", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 341, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.050822315737605095, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "6f1ad833-3d8a-4613-b5bd-a5d69263db1d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2107, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "6f1ad833-3d8a-4613-b5bd-a5d69263db1d", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "6f1ad833-3d8a-4613-b5bd-a5d69263db1d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2107, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.31179671082645655, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "e44ba4d3-b2d7-4a2a-b860-0fcbdda6c351", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1592, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "e44ba4d3-b2d7-4a2a-b860-0fcbdda6c351", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "e44ba4d3-b2d7-4a2a-b860-0fcbdda6c351", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1592, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.23224115278571844, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "fc7a2bcb-25d4-4840-a095-959499223267", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 139, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "fc7a2bcb-25d4-4840-a095-959499223267", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "fc7a2bcb-25d4-4840-a095-959499223267", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 139, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.021235859021544456, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "5d26ac5b-3c04-409d-a80d-5f1ac455916c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 139, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "5d26ac5b-3c04-409d-a80d-5f1ac455916c", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "5d26ac5b-3c04-409d-a80d-5f1ac455916c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 139, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.020852524787187576, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ed66c791-0393-45d5-946b-4cd5d3e8b395", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 116, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ed66c791-0393-45d5-946b-4cd5d3e8b395", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ed66c791-0393-45d5-946b-4cd5d3e8b395", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 116, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0182492071762681, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ba95d412-06d5-47d9-99b8-185146c2f869", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 185, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ba95d412-06d5-47d9-99b8-185146c2f869", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ba95d412-06d5-47d9-99b8-185146c2f869", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 185, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.027873259969055653, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "078a41ea-3a59-4b4f-b3c2-95cb6e538a6e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 81, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "078a41ea-3a59-4b4f-b3c2-95cb6e538a6e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "078a41ea-3a59-4b4f-b3c2-95cb6e538a6e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 81, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.013494960963726044, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d2b5a221-f890-4cdd-83bc-bdbff1f0d0b0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 82, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d2b5a221-f890-4cdd-83bc-bdbff1f0d0b0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d2b5a221-f890-4cdd-83bc-bdbff1f0d0b0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 82, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.012767246924340725, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "aaa378be-7365-4381-b2d3-35951ef88c7a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1800, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "aaa378be-7365-4381-b2d3-35951ef88c7a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "aaa378be-7365-4381-b2d3-35951ef88c7a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1800, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.2681455696001649, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "c2bb4f66-6955-4e0f-b223-b802c1cd2348", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1800, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "c2bb4f66-6955-4e0f-b223-b802c1cd2348", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "c2bb4f66-6955-4e0f-b223-b802c1cd2348", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1800, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.2688068374991417, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f2f01c2c-3327-4cae-a0d6-a12ec0e94fc3", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f2f01c2c-3327-4cae-a0d6-a12ec0e94fc3", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f2f01c2c-3327-4cae-a0d6-a12ec0e94fc3", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.006388836540281773, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "15856e8c-d105-478d-b0de-4ef8141b5a77", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 3028, "num_seqs": 5, "len_cu_seqlens": 6}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "15856e8c-d105-478d-b0de-4ef8141b5a77", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "15856e8c-d105-478d-b0de-4ef8141b5a77", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 3028, "num_seqs": 5, "len_cu_seqlens": 6}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.44989040307700634, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a39aa135-dd03-4abb-862d-020cf52fc7dd", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 61, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a39aa135-dd03-4abb-862d-020cf52fc7dd", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a39aa135-dd03-4abb-862d-020cf52fc7dd", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 61, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.010015377774834633, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "339a7ff4-0a55-40d0-99d1-f33e3b1ffe4a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 16, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "339a7ff4-0a55-40d0-99d1-f33e3b1ffe4a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "339a7ff4-0a55-40d0-99d1-f33e3b1ffe4a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 16, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.003360692411661148, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d8f4a9ae-d391-4f70-8069-f5deb510a2d1", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d8f4a9ae-d391-4f70-8069-f5deb510a2d1", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d8f4a9ae-d391-4f70-8069-f5deb510a2d1", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.006346044130623341, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d3dc3577-8a65-4937-a63a-4aa7204181a7", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d3dc3577-8a65-4937-a63a-4aa7204181a7", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d3dc3577-8a65-4937-a63a-4aa7204181a7", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.006266375072300434, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ce832e76-c4c2-421f-ad75-ebda2032c401", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 461, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ce832e76-c4c2-421f-ad75-ebda2032c401", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ce832e76-c4c2-421f-ad75-ebda2032c401", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 461, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.06927729770541191, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "6fbc155c-fcd0-4423-b1bf-970eef91d99b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 832, "num_seqs": 4, "len_cu_seqlens": 5}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "6fbc155c-fcd0-4423-b1bf-970eef91d99b", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "6fbc155c-fcd0-4423-b1bf-970eef91d99b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 832, "num_seqs": 4, "len_cu_seqlens": 5}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.12501337565481663, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a87ded8a-8202-405c-8ced-17521a8c8379", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2857, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a87ded8a-8202-405c-8ced-17521a8c8379", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a87ded8a-8202-405c-8ced-17521a8c8379", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2857, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.4273087615147233, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "62447caf-853a-4bf8-a62e-d12885248881", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 30, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "62447caf-853a-4bf8-a62e-d12885248881", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "62447caf-853a-4bf8-a62e-d12885248881", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 30, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.00546156894415617, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "fd072ba6-2190-4ce6-b96c-5212d4caf6a0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 114, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "fd072ba6-2190-4ce6-b96c-5212d4caf6a0", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "fd072ba6-2190-4ce6-b96c-5212d4caf6a0", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 114, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.018003075383603573, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "35ea9bbe-0189-4089-878c-4fc21461297e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 114, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "35ea9bbe-0189-4089-878c-4fc21461297e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "35ea9bbe-0189-4089-878c-4fc21461297e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 114, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.018130986019968987, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "1aa8cf18-0fca-4e32-a7ff-7ed29d24a3d2", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "1aa8cf18-0fca-4e32-a7ff-7ed29d24a3d2", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "1aa8cf18-0fca-4e32-a7ff-7ed29d24a3d2", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 35, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.006225933320820332, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d5f5c00c-c159-45c8-aa2a-2bb8a00e2fe8", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 31, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d5f5c00c-c159-45c8-aa2a-2bb8a00e2fe8", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d5f5c00c-c159-45c8-aa2a-2bb8a00e2fe8", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 31, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.005653194151818752, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "d5aa60dc-f159-4d02-a1d4-f68bd2a62f33", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 255, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "d5aa60dc-f159-4d02-a1d4-f68bd2a62f33", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "d5aa60dc-f159-4d02-a1d4-f68bd2a62f33", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 255, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.03795421030372381, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "28b70283-08fd-4eb9-8831-147abbf8d0f2", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 24, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "28b70283-08fd-4eb9-8831-147abbf8d0f2", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "28b70283-08fd-4eb9-8831-147abbf8d0f2", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 24, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.004441169090569019, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "73b8cc85-11ec-4ac4-8469-fce7b52ee44a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 97, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "73b8cc85-11ec-4ac4-8469-fce7b52ee44a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "73b8cc85-11ec-4ac4-8469-fce7b52ee44a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 97, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.015724758617579937, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "2683c087-e726-4a48-af25-674a95a889f2", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 97, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "2683c087-e726-4a48-af25-674a95a889f2", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "2683c087-e726-4a48-af25-674a95a889f2", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 97, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.015403139404952526, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "4b94d568-35ce-45a5-91eb-f8dc4b7077a7", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 574, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "4b94d568-35ce-45a5-91eb-f8dc4b7077a7", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "4b94d568-35ce-45a5-91eb-f8dc4b7077a7", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 574, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0857213269919157, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a0eb2dc2-9de3-47fc-8ab4-0beee6150015", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 92, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a0eb2dc2-9de3-47fc-8ab4-0beee6150015", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a0eb2dc2-9de3-47fc-8ab4-0beee6150015", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 92, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014635531231760979, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f3d30cb9-65d2-452a-92ef-10fb8b78473d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 85, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f3d30cb9-65d2-452a-92ef-10fb8b78473d", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f3d30cb9-65d2-452a-92ef-10fb8b78473d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 85, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.013616395182907581, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "7a7deca8-255f-4400-9fc6-367d4671ddcc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 79, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "7a7deca8-255f-4400-9fc6-367d4671ddcc", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "7a7deca8-255f-4400-9fc6-367d4671ddcc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 79, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.01273202896118164, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "977d19f8-5d0f-4c19-91e0-b05b500b34fc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 30, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "977d19f8-5d0f-4c19-91e0-b05b500b34fc", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "977d19f8-5d0f-4c19-91e0-b05b500b34fc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 30, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.00544558372348547, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "02c1e5f0-0d21-4c3c-a22d-40ad45a93549", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 30, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "02c1e5f0-0d21-4c3c-a22d-40ad45a93549", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "02c1e5f0-0d21-4c3c-a22d-40ad45a93549", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 30, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.005356903187930584, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "5a91aa02-8cee-49f9-aa2b-145d6e254ee4", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 401, "num_seqs": 4, "len_cu_seqlens": 5}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "5a91aa02-8cee-49f9-aa2b-145d6e254ee4", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "5a91aa02-8cee-49f9-aa2b-145d6e254ee4", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 401, "num_seqs": 4, "len_cu_seqlens": 5}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.05984139163047075, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "8e7ef744-5f4e-40e7-aae3-471e013657bd", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 12, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "8e7ef744-5f4e-40e7-aae3-471e013657bd", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "8e7ef744-5f4e-40e7-aae3-471e013657bd", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 12, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0027241799980401993, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "85d7becb-2969-4cc8-9120-1ec26a78d258", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 18, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "85d7becb-2969-4cc8-9120-1ec26a78d258", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "85d7becb-2969-4cc8-9120-1ec26a78d258", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 18, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0036588450893759727, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "e286a4f4-4a93-43bd-9465-b2915f0f8f96", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1796, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "e286a4f4-4a93-43bd-9465-b2915f0f8f96", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "e286a4f4-4a93-43bd-9465-b2915f0f8f96", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 1796, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.26614286098629236, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "08d4f2c4-dd07-4cc0-9d93-975c91b17577", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 973, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "08d4f2c4-dd07-4cc0-9d93-975c91b17577", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "08d4f2c4-dd07-4cc0-9d93-975c91b17577", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 973, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.14641222450882196, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "9c1ef562-578a-4df4-986e-e0c1aaa4727a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 973, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "9c1ef562-578a-4df4-986e-e0c1aaa4727a", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "9c1ef562-578a-4df4-986e-e0c1aaa4727a", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 973, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.1460703583434224, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "bfd8f7b6-3953-46c3-af04-409b2fbe353c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 248, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "bfd8f7b6-3953-46c3-af04-409b2fbe353c", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "bfd8f7b6-3953-46c3-af04-409b2fbe353c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 248, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.037331245839595795, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "c358edcd-3fda-43bb-98a2-ac32da9038b6", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2040, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "c358edcd-3fda-43bb-98a2-ac32da9038b6", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "c358edcd-3fda-43bb-98a2-ac32da9038b6", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2040, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.3021991178393364, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f203fdcd-6b72-4687-9fcf-a010066eab5c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 48, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f203fdcd-6b72-4687-9fcf-a010066eab5c", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f203fdcd-6b72-4687-9fcf-a010066eab5c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 48, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.008138821460306644, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "33a38713-868b-43f7-8d75-fbe472e30387", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 118, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "33a38713-868b-43f7-8d75-fbe472e30387", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "33a38713-868b-43f7-8d75-fbe472e30387", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 118, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.01859570387750864, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "3a77dfec-f1c1-40aa-82e4-fafdfc84e540", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 76, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "3a77dfec-f1c1-40aa-82e4-fafdfc84e540", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "3a77dfec-f1c1-40aa-82e4-fafdfc84e540", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 76, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.014340773224830627, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "ea27be17-3dfa-423e-aa85-e9001c24daed", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 76, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "ea27be17-3dfa-423e-aa85-e9001c24daed", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "ea27be17-3dfa-423e-aa85-e9001c24daed", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 76, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.012340988032519817, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "49ef89d2-da00-4cd5-b912-708cbfc7106d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 67, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "49ef89d2-da00-4cd5-b912-708cbfc7106d", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "49ef89d2-da00-4cd5-b912-708cbfc7106d", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 67, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.010950461961328983, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "056224b8-0a27-4634-b93b-36beb658e135", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 68, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "056224b8-0a27-4634-b93b-36beb658e135", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "056224b8-0a27-4634-b93b-36beb658e135", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 68, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.011172923259437084, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "685d26ff-9ad1-4378-b15a-b279d2db1bcb", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 13, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "685d26ff-9ad1-4378-b15a-b279d2db1bcb", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "685d26ff-9ad1-4378-b15a-b279d2db1bcb", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 13, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0028944388031959534, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "352c9ace-2650-481f-ab51-69394fd0f366", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 525, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "352c9ace-2650-481f-ab51-69394fd0f366", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "352c9ace-2650-481f-ab51-69394fd0f366", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 525, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.07852388825267553, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "2c9693b4-6375-459a-a29f-650c24abe472", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 49, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "2c9693b4-6375-459a-a29f-650c24abe472", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "2c9693b4-6375-459a-a29f-650c24abe472", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 49, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.008299603126943111, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "27f44fd6-0b0d-44bc-a771-617fd3e9e61c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 49, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "27f44fd6-0b0d-44bc-a771-617fd3e9e61c", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "27f44fd6-0b0d-44bc-a771-617fd3e9e61c", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 49, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.008306935429573059, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "07aa7922-1848-48a9-830a-54216b5553b3", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 5709, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "07aa7922-1848-48a9-830a-54216b5553b3", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "07aa7922-1848-48a9-830a-54216b5553b3", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 5709, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.8496198868378997, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "eaa0fd47-ac57-47d0-a896-e3b1a914a5da", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 14, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "eaa0fd47-ac57-47d0-a896-e3b1a914a5da", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "eaa0fd47-ac57-47d0-a896-e3b1a914a5da", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 14, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0030686743557453156, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "f105eda8-d047-4624-90a2-f5d5039e2f4e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 661, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "f105eda8-d047-4624-90a2-f5d5039e2f4e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "f105eda8-d047-4624-90a2-f5d5039e2f4e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 661, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.09857802093029022, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "cd979341-7112-4887-a2f3-dfb269a49bbc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 76, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "cd979341-7112-4887-a2f3-dfb269a49bbc", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "cd979341-7112-4887-a2f3-dfb269a49bbc", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 76, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.012277862057089806, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "43bf9699-adb3-4161-b741-8ff07b553336", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 959, "num_seqs": 4, "len_cu_seqlens": 5}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "43bf9699-adb3-4161-b741-8ff07b553336", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "43bf9699-adb3-4161-b741-8ff07b553336", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 959, "num_seqs": 4, "len_cu_seqlens": 5}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.14233729429543018, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "54856fec-3b50-4866-815d-19bfa16cf58b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 959, "num_seqs": 4, "len_cu_seqlens": 5}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "54856fec-3b50-4866-815d-19bfa16cf58b", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "54856fec-3b50-4866-815d-19bfa16cf58b", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 959, "num_seqs": 4, "len_cu_seqlens": 5}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.14402965921908617, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "2ba465c0-7405-4bf1-ba4e-7d632d95f141", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 28, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "2ba465c0-7405-4bf1-ba4e-7d632d95f141", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "2ba465c0-7405-4bf1-ba4e-7d632d95f141", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 28, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.005222685635089874, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "1efaf2a9-05db-4737-8f41-6880ba1bb487", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 42, "num_seqs": 2, "len_cu_seqlens": 3}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "1efaf2a9-05db-4737-8f41-6880ba1bb487", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "1efaf2a9-05db-4737-8f41-6880ba1bb487", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 42, "num_seqs": 2, "len_cu_seqlens": 3}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.0072821686044335365, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "a01a3f93-19b0-4b12-95e3-6132a41d628e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2284, "num_seqs": 3, "len_cu_seqlens": 4}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "a01a3f93-19b0-4b12-95e3-6132a41d628e", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "a01a3f93-19b0-4b12-95e3-6132a41d628e", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 2284, "num_seqs": 3, "len_cu_seqlens": 4}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.34308723639696836, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}
{"uuid": "cc241d2e-f8b7-448f-95ec-1d5d2809c068", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 23, "num_seqs": 1, "len_cu_seqlens": 2}, "source": "https://huggingface.co/datasets/flashinfer-ai/mlsys26-contest/tree/5e832ce88dd1013032b22a0e942979d7d2769cc3", "source_workload": "cc241d2e-f8b7-448f-95ec-1d5d2809c068", "model": "Qwen3-Next (TP=4), as recorded by the MLSys 2026 FlashInfer contest", "adapter": "Imported unchanged from the MLSys 2026 FlashInfer contest dataset; each input tensor is stored in its own content-addressed file. The contest publishes no engine outputs.", "validation": {"uuid": "cc241d2e-f8b7-448f-95ec-1d5d2809c068", "definition": "gdn_prefill_qk4_v8_d128_k_last", "axes": {"total_seq_len": 23, "num_seqs": 1, "len_cu_seqlens": 2}, "status": "passed", "device": "cuda", "torch": "2.14.1+cu130", "seconds": 0.004298948682844639, "comparison": "reference outputs have the declared shapes and dtypes and are finite; the contest publishes no engine outputs to compare against"}}