diff --git "a/captures.jsonl" "b/captures.jsonl"
--- "a/captures.jsonl"
+++ "b/captures.jsonl"
@@ -1,338 +1,338 @@
 {"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, "Nall": 4121, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 25, "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"}, "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, "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"}}
+{"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, "Nall": 4121, "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"}, "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, "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"}}
+{"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, "Nall": 4121, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 30, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 4125, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 32, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
+{"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, "Nall": 4121, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 25, "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"}, "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, "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"}}
+{"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, "Nall": 4121, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 30, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 4125, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 32, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
+{"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, "Nall": 4121, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 25, "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"}, "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, "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"}}
+{"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, "Nall": 4121, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 30, "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"}, "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, "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"}}
+{"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, "Nall": 6174, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 4125, "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"}, "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, "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"}}
+{"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, "Nall": 8221, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
-{"adapter": "Preserve post-normalization/RoPE tensors; materialize the engine's segment visibility.", "axes": {"B": 1, "Nall": 32, "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"}, "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, "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"}}
+{"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, "Nall": 4128, "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"}, "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, "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"}}
+{"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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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_d128/2ac2a1d1e80676c131af9c5b87b012ab989b4d17087129af1e3efc364a8aee91.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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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_d128/6307cb0a22b7ac2e920fe9ad9453d89ef42f0b815f7cc695fe30c48ec9ad0395.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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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/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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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/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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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/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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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/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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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/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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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_d128/73659cd428a2c0d0fb25ae46514c9697d1ac106f6c76e6d411054e060ac19c72.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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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"}, "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", "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"}, "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/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"}