{ "format": "agentx-fpm-submission-inputs-v1", "dataset_base_revision": "664e813143ab3f52608fd6bb4b11a7c2d838dc34", "reproduction_repository": "https://github.com/ai-dynamo/aisimulate", "reproduction_revision": "037c0aea953960187aeb07f4ae656614ed7585e6", "aiperf_repository": "https://github.com/SemiAnalysisAI/aiperf", "aiperf_revision": "754356e9a39acc6cc6afb242d123bb57c3fb6f75", "cases": [ { "case_id": "glm-5.2-440958-b200-sglang-job4208956", "job_id": 4208956, "reference_id": 440958, "source_capture": "agentx-fpm-ab-20260909/job-4208956/on/fpm.jsonl", "source_sha256": "0c6e52190021247e095f518c8413a4d8afc2b9c38aab8bbe67fd1ab0a9e1c8bf", "source_export": "agentx-fpm-ab-20260909/job-4208956/on/aiperf/profile_export_aiperf.json", "export_timezone": "America/Los_Angeles", "expected_dp_ranks": [0], "model_id": "nvidia/GLM-5.2-NVFP4", "model_revision": "53e0691e21895a3863a606dfd12910c69eba94ab", "hardware": "8x B200, umb-b200-041, Computelab Slurm", "backend": "sglang", "base_source_revision": "71de97b264b04dcd514cf904003028aefe9775c8", "fpm_patch_revision": "ed18d64951b93ac252d921ee037ce0d3327eda1f", "image": "nvcr.io/nvidian/dynamo-dev/sglang-agentx@sha256:f856a45537f82e1900ea7607edcbaa7f77fbb2e70220eae522d1d50d0046727e", "parallelism": {"attention_tp": 8, "attention_dp": 1, "pp": 1, "moe_tp": 8, "moe_ep": 1}, "concurrency": 4, "g2": {"enabled": false, "hicache": false}, "speculation": {"method": "EAGLE/native MTP", "steps": 3, "topk": 1, "draft_tokens": 4, "synthetic_acceptance_length": 2.99, "synthetic_method": "match-expected", "synthetic_token_mode": "real-draft-token"}, "serving": {"frontend": "Dynamo frontend + dynamo.sglang", "kv_cache_dtype": "fp8_e4m3", "quantization": "modelopt_fp4", "chunked_prefill_size": 8192, "max_prefill_tokens": 8192, "max_running_requests": 8, "cuda_graph_max_bs": 8, "mem_fraction_static": 0.83}, "timing_scope": "SGLang existing CUDA DeviceTimer boundaries with PR38711 speculative segment timing fix; not client latency", "recorder": {"component": "Dynamo Python disk recorder", "revision": "1926d6ec765fcf7e86fc000a1b6a15b2b107e607", "transport": "Dynamo FPM subscription, buffered JSONL, one-second flush", "clock": "time.time_ns() on one recorder"}, "reproduction_directory": "docs/experiments/agentx-glm-5.2-440958-b200-sglang-slurm", "launcher": "run-fpm.sh", "known_caveats": ["Historical off/on runs used different nodes; no causal overhead claim.", "Naive AIPerf timestamps are PDT, verified against epoch-nanosecond profiling request timestamps.", "Affected Prometheus counter deltas are not used as measurement evidence."] }, { "case_id": "deepseek-v4-pro-440845-b300-sglang-job4209414", "job_id": 4209414, "reference_id": 440845, "source_capture": "agentx-dsv4-440845-ab-20260909/analysis-4209414/fpm.jsonl", "source_sha256": "bf0f277dbd74a8276e0ed8021a882a199a0a394f18132fded80c3e6fa2df2614", "source_export": "agentx-dsv4-440845-ab-20260909/job-4209414/on/aiperf/profile_export_aiperf.json", "export_timezone": "UTC", "expected_dp_ranks": [0, 1, 2, 3, 4, 5, 6, 7], "model_id": "deepseek-ai/DeepSeek-V4-Pro", "model_revision": "b5968e9190ef611bbf34a7229255be88a0e937c1", "hardware": "8x B300, umb-b300-dp-127, Computelab Slurm", "backend": "sglang", "base_source_revision": "71de97b264b04dcd514cf904003028aefe9775c8", "fpm_patch_revision": "ed18d64951b93ac252d921ee037ce0d3327eda1f", "image": "nvcr.io/nvidian/dynamo-dev/sglang-agentx@sha256:f856a45537f82e1900ea7607edcbaa7f77fbb2e70220eae522d1d50d0046727e", "parallelism": {"attention_tp": 1, "attention_dp": 8, "pp": 1, "moe_tp": 1, "moe_ep": 8, "sglang_cli_tp": 8}, "concurrency": 32, "g2": {"enabled": false, "hicache": false}, "speculation": {"method": "EAGLE/native MTP", "steps": 3, "topk": 1, "draft_tokens": 4, "synthetic_acceptance_length": 2.49, "synthetic_method": "match-expected", "synthetic_token_mode": "real-draft-token"}, "serving": {"frontend": "native SGLang + sglang-router consistent hashing, DP-aware", "kv_cache_dtype": "fp8_e4m3", "attention_backend": "dsv4", "moe_backend": "megamoe", "chunked_prefill_size_cli_global": 65536, "chunked_prefill_size_per_rank": 8192, "max_running_requests_cli": 64, "cuda_graph_max_bs_decode": 544, "mem_fraction_static": 0.93, "stream_interval": 20}, "timing_scope": "SGLang existing CUDA DeviceTimer boundaries with PR38711 speculative segment timing fix; positive-time empty-rank work retained", "recorder": {"component": "record_fpm.py", "transport": "8 rank-local IPC ZMQ subscriptions in one recorder, buffered JSONL", "clock": "time.time_ns() on one recorder"}, "reproduction_directory": "docs/experiments/agentx-deepseek-v4-pro-440845-b300-sglang-slurm", "launcher": "campaign.py", "known_caveats": ["One ordered off/on pair; not a statistical zero-overhead bound.", "AIPerf forkserver cleanup was recovered before each measured phase; measured configuration was unchanged.", "This fixed run uses full-capture positive-time ordinal alignment after equal-count, continuous-counter, global-correlation and windowed lag audits; this is inferred, not a publisher-certified global ID.", "SGLang --tp 8 with attention DP8 means attention TP1, not 64 GPUs."] }, { "case_id": "deepseek-v4-pro-440246-b200-vllm-job4228930", "job_id": 4228930, "reference_id": 440246, "source_capture": "vllm-fpm-sd-20260910/job-4228930/on/fpm.jsonl", "source_sha256": "7c57d51ec58e66917ee54ddddf7983c7e126cf3f6e7b5a6b57feb9fdf1d623d8", "source_export": "vllm-fpm-sd-20260910/job-4228930/on/aiperf/profile_export_aiperf.json", "export_timezone": "UTC", "expected_dp_ranks": [0, 1, 2, 3, 4, 5, 6, 7], "model_id": "deepseek-ai/DeepSeek-V4-Pro", "model_revision": "b5968e9190ef611bbf34a7229255be88a0e937c1", "hardware": "8x B200, Computelab Slurm exclusive node", "backend": "vllm", "backend_version": "0.28.0", "base_source_revision": "2cf0a6915ce544dc493a0990f2ea38d81601128a", "fpm_patch_revision": "996fed467139edd7719a0063d57709b8a7fa6989", "image": "nvcr.io/nvidian/dynamo-dev/vllm-agentx@sha256:b58174af2daaf4d02c4845fc90dd3f18615fc7c3ef930fb852ed0d29e76f1d12", "parallelism": {"attention_tp": 1, "attention_dp": 8, "pp": 1, "moe_tp": 1, "moe_ep": 8}, "concurrency": 64, "g2": {"enabled": true, "connector": "SimpleCPUOffloadConnector", "bytes_per_rank": 137438953472, "total_bytes": 1099511627776}, "speculation": {"method": "mtp", "speculative_tokens": 3, "synthetic_acceptance_length": 2.49}, "serving": {"frontend": "native vLLM Rust frontend + vllm-router 0.1.14", "scheduler": "native AsyncScheduler, not Dynamo InstrumentedScheduler", "model_runner": "V2", "kv_cache_dtype": "fp8", "page_size": 256, "max_model_len": 1048576, "max_num_seqs_per_rank": 16, "max_num_batched_tokens": 8192, "prefill_schedule_interval": 8, "gpu_memory_utilization": 0.9, "plugins": "VLLM_PLUGINS empty; text-only", "numa_binding": "explicit PCI-to-NUMA mapping"}, "timing_scope": "model_step_cuda", "recorder": {"component": "record_fpm.py", "transport": "8 native vLLM TCP ZMQ subscriptions in one recorder, buffered JSONL", "clock": "time.time_ns() on one recorder"}, "reproduction_directory": "docs/experiments/agentx-deepseek-v4-pro-440246-b200-vllm", "launcher": "benchmark.py", "known_caveats": ["Native CUDA timing spans model execution, sampling and speculative drafting; it is not E2E latency or pure GPU kernel busy time.", "Empty-rank dummy steps are not emitted; do not fabricate missing rank records.", "Independent async publication can shift receive-time correspondence by one iteration.", "FPM-on profiling had 23 drain cancellations; their pre-cutoff GPU work is retained."] }, { "case_id": "minimax-m3-439922-b200-vllm-job4245705", "job_id": 4245705, "reference_id": 439922, "source_capture": "minimax-m3-439922-vllm-20260911/job-4245705/on/fpm.jsonl", "source_sha256": "18d3fcf012b51aae745296853d0aebee99288b9bc6244828407c462b76401a19", "source_export": "minimax-m3-439922-vllm-20260911/job-4245705/on/aiperf/profile_export_aiperf.json", "export_timezone": "UTC", "expected_dp_ranks": [0], "model_id": "nvidia/MiniMax-M3-NVFP4", "model_revision": "901464083161bf8612a29ff7ad29914cd4ab4a85", "hardware": "4x B200, umbriel-b200-048, Computelab Slurm, partial-node allocation", "backend": "vllm", "backend_version": "0.28.0", "base_source_revision": "2cf0a6915ce544dc493a0990f2ea38d81601128a", "fpm_patch_revision": "b3563fc65ae0f5359802593d78e7ea097e1fed31", "image": "nvcr.io/nvidian/dynamo-dev/vllm-agentx@sha256:2aab1ad231c052eacc53f430b37261b01b31abc1319cd72e0e5cf9936292199a", "parallelism": {"attention_tp": 4, "attention_dp": 1, "pp": 1, "moe_tp": 4, "moe_ep": 1}, "concurrency": 15, "g2": {"enabled": false, "cpu_offload_gb": 0, "kv_connector": null}, "speculation": {"method": "eagle3", "draft_model": "Inferact/MiniMax-M3-EAGLE3-GQA", "draft_revision": "96692486b5fd38ebf8fd2a5f6bb53427d30819a8", "speculative_tokens": 3, "synthetic_acceptance_length": 2.78}, "serving": {"frontend": "native vLLM Python", "scheduler": "native vLLM", "model_runner": "V1", "kv_cache_dtype": "fp8", "page_size": 128, "max_model_len": 1048576, "max_num_batched_tokens": 16384, "stream_interval": 20, "gpu_memory_utilization": 0.9}, "timing_scope": "model_step_cuda", "recorder": {"component": "record_fpm.py inherited from native vLLM campaign", "transport": "one native TCP ZMQ subscription, buffered JSONL", "clock": "time.time_ns() on one recorder"}, "reproduction_directory": "docs/experiments/agentx-minimax-m3-439922-b200-vllm", "launcher": "benchmark.py", "known_caveats": ["Image-only no-copy FlashAttention descale layout fix, in addition to the unchanged FPM patch.", "One measured Broken pipe request error; no profiling cancellations. The capture retains work before cutoff, not only successful requests.", "FPM off/on used separate partial-node allocations; not a controlled micro-overhead comparison.", "Native FPM field semantics are preserved by producer revision; no variance or timing re-interpretation is performed."] } ] }