{
  "label": "tp4-mtp-k1-replayssm-iter19-graph-capsmall-full",
  "stamp": "20260622222755",
  "return_codes": {
    "metrics": 0,
    "json_canary": 1,
    "color_canary": 1,
    "quality_suite": 0
  },
  "status": {
    "metrics": "passed",
    "json_canary": "failed",
    "color_canary": "failed",
    "quality_suite": "skipped"
  },
  "artifacts": {
    "metrics": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full-p512o512-20260622222755.json",
    "json_canary": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full-json-repeat96-20260622222755.json",
    "color_canary": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full-color-repeat96-20260622222755.json",
    "quality_suite": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full-quality-suite-20260622222755.json",
    "log": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full-20260622222755.log"
  },
  "artifact_exists": {
    "metrics": true,
    "json_canary": true,
    "color_canary": true,
    "quality_suite": false,
    "log": true
  },
  "env": {
    "MODEL_PATH": "/mnt/fast-ai/llm-cache/hf/models--nameistoken--Qwen3.6-35B-A3B-Quark-W8A8-INT8/snapshots/cced56592e8c8935f8220836b4baa04dfd389118",
    "SERVED_MODEL_NAME": "qwen36-35b-a3b-fp8",
    "SERVER_LAUNCHER": "/home/steve/llm-optimizations/scripts/launch-qwen36-quark-int8-hybrid-mtp.sh",
    "TP_SIZE": "4",
    "MAX_MODEL_LEN": "32768",
    "MAX_NUM_BATCHED_TOKENS": "8192",
    "MAX_NUM_SEQS": "48",
    "ONEAPI_DEVICE_SELECTOR": "level_zero:0,1,2,3",
    "ZE_AFFINITY_MASK": "0,1,2,3",
    "QWEN36_XPU_PREFLIGHT": "auto",
    "CACHE_LABEL": "qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full",
    "TORCHINDUCTOR_CACHE_DIR": "/mnt/fast-ai/vllm-cache-exp/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full/torchinductor",
    "VLLM_CACHE_ROOT": "/mnt/fast-ai/vllm-cache-exp/qwen36-ablation-tp4-mtp-k1-replayssm-iter19-graph-capsmall-full/vllm",
    "BASE_URL": "http://127.0.0.1:18080",
    "PORT": "18080",
    "VLLM_XPU_GDN_NATIVE_FALLBACK": "decode,prefill",
    "VLLM_XPU_GDN_NATIVE_PREFILL_FALLBACK": null,
    "VLLM_XPU_GDN_PREFILL_RECURRENT_FALLBACK": "1",
    "VLLM_XPU_GDN_NATIVE_FALLBACK_INITIAL_DECODE_STEPS": null,
    "VLLM_XPU_GDN_NATIVE_FALLBACK_DECODE_LAYERS": null,
    "VLLM_XPU_GDN_NATIVE_FALLBACK_DECODE_MAX_LAYER": null,
    "VLLM_XPU_GDN_NATIVE_FALLBACK_DECODE_LAYER_REGEX": null,
    "VLLM_XPU_GDN_SYNC_BEFORE_NATIVE_DECODE": null,
    "VLLM_XPU_GDN_SYNC_BEFORE_NATIVE_DECODE_LAYERS": null,
    "VLLM_XPU_GREEDY_SAMPLE_TOPK_FALLBACK": null,
    "VLLM_XPU_GREEDY_SAMPLE_XPU_FALLBACK": null,
    "VLLM_XPU_GREEDY_SAMPLE_BF16_TOPK_FASTPATH": null,
    "VLLM_XPU_ASYNC_CLONE_SAMPLED_TOKEN_IDS": null,
    "VLLM_XPU_SANITIZE_LOGITS_NAN": null,
    "VLLM_XPU_LOCAL_ARGMAX_DECODE": null,
    "VLLM_XPU_LOCAL_ARGMAX_DIRECT_GATHER": null,
    "VLLM_XPU_LOCAL_ARGMAX_DIRECT_GATHER_REUSE": null,
    "VLLM_XPU_LOCAL_ARGMAX_PACKED_GATHER": null,
    "VLLM_XPU_LOCAL_ARGMAX_SYNC_TOKENS": null,
    "VLLM_XPU_DUMMY_LOGIT_INDICES_STATIC": null,
    "VLLM_XPU_ZERO_ALL_PREFILL_GDN_STATE": null,
    "VLLM_XPU_ZERO_FRESH_GDN_STATE": null,
    "VLLM_XPU_DISABLE_PREFILL_CUDAGRAPH_REPLAY": "0",
    "VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE": null,
    "XPU_GRAPH": "1",
    "VLLM_XPU_ENABLE_XPU_GRAPH": "1",
    "VLLM_XPU_FORCE_GRAPH_WITH_COMM": "1",
    "VLLM_XPU_GRAPH_NOOP_COMM_CAPTURE": "1",
    "VLLM_XPU_USE_CUSTOM_OP_COLLECTIVES": null,
    "VLLM_XPU_COMPILE_ALLREDUCE_CUSTOM_OP": null,
    "VLLM_XPU_CUSTOM_ALLREDUCE_GRAPH_CLONE_INPUT": null,
    "VLLM_XPU_CUSTOM_ALLREDUCE_CLONE_INPUT": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_MAX_PIECEWISE_INDEX": null,
    "VLLM_XPU_CUDAGRAPH_DISABLE_SUBMOD_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_RECAPTURE_AFTER_N_REPLAYS": null,
    "VLLM_XPU_CUDAGRAPH_RECAPTURE_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_STRONG_OUTPUT": null,
    "VLLM_XPU_CUDAGRAPH_PER_WRAPPER_POOL": null,
    "VLLM_XPU_CUDAGRAPH_NO_GLOBAL_POOL": null,
    "VLLM_XPU_DECODE_CUDAGRAPH_REPLAY_EAGER_EVERY_N_REQUESTS": null,
    "VLLM_XPU_CUDAGRAPH_ALLOW_RUNTIME_RECAPTURE": null,
    "VLLM_XPU_SYNC_CUDAGRAPH_REPLAY": null,
    "VLLM_XPU_CUDAGRAPH_SANITIZE_REPLAY_INPUTS": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_FILE": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_MAX_LINES": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_RANK": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_SUBMOD_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_REQ_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_INPUTS": null,
    "VLLM_XPU_CUDAGRAPH_REPLAY_TRACE_DIGEST": null,
    "VLLM_XPU_GDN_TRACE_FILE": null,
    "VLLM_XPU_GDN_TRACE_MAX_LINES": null,
    "VLLM_XPU_GDN_TRACE_RANK": null,
    "VLLM_XPU_GDN_TRACE_LAYER_REGEX": null,
    "VLLM_XPU_GDN_TRACE_REQ_REGEX": null,
    "VLLM_XPU_GDN_TRACE_PREFILL_ONLY": null,
    "VLLM_XPU_GDN_TRACE_DECODE_ONLY": null,
    "VLLM_XPU_GDN_TRACE_TENSOR_LIMIT": null,
    "VLLM_XPU_GDN_TRACE_STATE_LIMIT": null,
    "VLLM_XPU_CUDAGRAPH_ZERO_REPLAY_OUTPUT_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_ZERO_REPLAY_OUTPUT_INDICES": null,
    "VLLM_XPU_CUDAGRAPH_CLONE_REPLAY_OUTPUT_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_CLONE_REPLAY_OUTPUT_INDICES": null,
    "VLLM_XPU_CUDAGRAPH_STATIC_INPUT_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_STATIC_INPUT_INDICES": null,
    "VLLM_XPU_CUDAGRAPH_STATIC_INPUT_MAX_NUMEL": null,
    "VLLM_XPU_CUDAGRAPH_COMPARE_DIRECT_REGEX": null,
    "VLLM_XPU_CUDAGRAPH_COMPARE_DIRECT_RETURN_DIRECT": null,
    "VLLM_XPU_W8A8_EXPERIMENTAL_ALLOW": null,
    "VLLM_XPU_W8A8_USE_OFFSETS": null,
    "VLLM_XPU_W8A8_OFFSETS_PREFIX_OP": null,
    "VLLM_XPU_MOE_W8A8_MIDDLE_LAYERLET": null,
    "VLLM_XPU_MOE_W8A8_FULL_LAYERLET": null,
    "VLLM_XPU_MOE_W8A8_FUSED_Q1": null,
    "VLLM_XPU_MOE_W8A8_FAST_GATHER": null,
    "VLLM_XPU_MOE_W8A8_DIRECT_GEMM2_GATHER": null,
    "VLLM_XPU_MOE_W8A8_DPAS_GEMM2_GATHER": null,
    "VLLM_XPU_MOE_W8A8_DPAS_GEMM2_GATHER_NTILE": null,
    "VLLM_XPU_MOE_W8A8_WORKSPACE_ATOMIC": null,
    "VLLM_XPU_MOE_W8A8_ROUTE_GEMM1": null,
    "VLLM_XPU_MOE_W8A8_ROUTE_GEMM1_MTILE": null,
    "VLLM_XPU_INT8_MOE_MIXED_WORKSPACE": null,
    "VLLM_XPU_INT8_MOE_PERSISTENT_SCRATCH": null,
    "VLLM_XPU_INT8_MOE_FUSED_PROLOGUE_OFFSET": null,
    "VLLM_XPU_INT8_MOE_FUSED_PROLOGUE_OFFSET_ALLOW_CAPTURE": null,
    "VLLM_XPU_INT8_MOE_FUSED_PROLOGUE_OFFSET_SKIP_LAYERS": null,
    "VLLM_XPU_INT8_MOE_FUSED_PROLOGUE_OFFSET_LAYER_REGEX": null,
    "VLLM_XPU_INT8_MOE_ACTIVE_OFFSET_GEMM": null,
    "VLLM_XPU_SHARED_EXPERTS_STREAM": null,
    "VLLM_XPU_SHARED_EXPERT_FUSED_ACT_QUANT": null,
    "VLLM_XPU_SHARED_EXPERT_FUSED_ACT_QUANT_OUT": null,
    "VLLM_XPU_SHARED_EXPERT_BOUNDARY_OUT": null,
    "VLLM_XPU_SHARED_EXPERT_FUSED_GATE_MUL": null,
    "VLLM_XPU_MOE_SHARED_ADD_ALLREDUCE_CUSTOM_OP": null,
    "VLLM_XPU_METADATA_COPY_ALLOW": null,
    "VLLM_XPU_BLOCK_TABLE_DIRTY_COMMIT": null,
    "VLLM_XPU_BLOCK_TABLE_DIRTY_COMMIT_LOG_EVERY": null,
    "GPU_MEMORY_UTILIZATION": "0.95",
    "COMPILATION_CONFIG": "{\"cudagraph_mode\":\"PIECEWISE\",\"max_cudagraph_capture_size\":4}",
    "SPEC_CONFIG": null,
    "ABLATION_FAST_GRAPH_AUTOCONFIG": "1",
    "VLLM_EXTRA_ARGS": "--no-async-scheduling --mamba-cache-mode align",
    "METRICS_PROMPT_KIND": null,
    "METRICS_PROMPT_PRESET": null,
    "METRICS_RANDOM_PREFIX_LEN": null,
    "METRICS_PROMPT_TOKENS": null,
    "METRICS_OUTPUT_TOKENS": null,
    "METRICS_WARMUP_OUTPUT_TOKENS": null,
    "JSON_LOGPROBS": null,
    "COLOR_LOGPROBS": null
  },
  "metrics_summary": {
    "tok_s_out_client_after_first_chunk": {
      "mean": 76.3937572446695,
      "median": 76.3937572446695,
      "min": 75.98526305057288,
      "max": 76.8022514387661
    },
    "tok_s_out_client_after_first_chunk_corrected": {
      "mean": 76.244550687551,
      "median": 76.244550687551,
      "min": 75.83685433367722,
      "max": 76.65224704142477
    },
    "tok_s_out_client_e2e": {
      "mean": 74.26534305480835,
      "median": 74.26534305480835,
      "min": 73.89072194456203,
      "max": 74.63996416505468
    },
    "tok_s_total_client": {
      "mean": 146.4999931354618,
      "median": 146.4999931354618,
      "min": 145.76099446095245,
      "max": 147.23899180997114
    },
    "ttft_ms_client": {
      "mean": 192.06374849454733,
      "median": 192.06374849454733,
      "min": 191.00272699142806,
      "max": 193.1247699976666
    },
    "ttft_ms_vllm_metrics": {
      "mean": 177.30772495269775,
      "median": 177.30772495269775,
      "min": 176.17154121398926,
      "max": 178.44390869140625
    },
    "tok_s_prefill_lower_bound_from_ttft": {
      "mean": 2592.968232599591,
      "median": 2592.968232599591,
      "min": 2578.6438477364504,
      "max": 2607.2926174627314
    },
    "e2e_ms_vllm_metrics": {
      "mean": 6893.291234970093,
      "median": 6893.291234970093,
      "min": 6858.558177947998,
      "max": 6928.0242919921875
    },
    "queue_ms_vllm_histogram": {
      "mean": 0.012458498531486839,
      "median": 0.012458498531486839,
      "min": 0.011801996151916683,
      "max": 0.013115000911056995
    },
    "prefill_ms_vllm_histogram": {
      "mean": 172.11763299565064,
      "median": 172.11763299565064,
      "min": 170.71417499391828,
      "max": 173.521090997383
    },
    "decode_ms_vllm_histogram": {
      "mean": 6716.240793000907,
      "median": 6716.240793000907,
      "min": 6680.353828996886,
      "max": 6752.127757004928
    },
    "decode_ms_per_generation_token_vllm_histogram": {
      "mean": 13.117657798829896,
      "median": 13.117657798829896,
      "min": 13.047566072259542,
      "max": 13.18774952540025
    },
    "inference_ms_vllm_histogram": {
      "mean": 6888.358425996557,
      "median": 6888.358425996557,
      "min": 6853.874919994269,
      "max": 6922.841931998846
    },
    "time_per_output_token_ms_vllm_histogram": {
      "mean": 13.143328362037007,
      "median": 13.143328362037007,
      "min": 13.073099469661232,
      "max": 13.213557254412784
    },
    "inter_token_ms_vllm_histogram": {
      "mean": 13.143328362037,
      "median": 13.143328362037,
      "min": 13.073099469661225,
      "max": 13.213557254412775
    },
    "iteration_tokens_per_step_vllm_histogram": {
      "mean": 1.97265625,
      "median": 1.97265625,
      "min": 1.97265625,
      "max": 1.97265625
    }
  },
  "speed": {
    "tok_s_out_corrected_mean": 76.244550687551,
    "tok_s_out_after_first_chunk_mean": 76.3937572446695,
    "tok_s_out_e2e_mean": 74.26534305480835,
    "decode_ms_per_token_mean": 13.117657798829896,
    "ttft_ms_client_mean": 192.06374849454733
  },
  "json_canary": {
    "pass_all": false,
    "repeats_completed": 83,
    "mismatch_count": 1,
    "first_mismatch": {
      "index": 82,
      "request_id": null,
      "text": "{\"key\": \"12\", \"unit\": \"widgets\"}",
      "normalized": "{\"key\": \"12\", \"unit\": \"widgets\"}",
      "token_ids": [
        4754,
        767,
        763,
        328,
        16,
        17,
        487,
        328,
        3715,
        763,
        328,
        28844,
        8934
      ],
      "finish_reason": "stop",
      "elapsed_s": 0.2745149819966173,
      "usage": {
        "prompt_tokens": 36,
        "total_tokens": 50,
        "completion_tokens": 14,
        "prompt_tokens_details": null
      },
      "json_parsed": {
        "key": "12",
        "unit": "widgets"
      },
      "pass": false
    }
  },
  "color_canary": {
    "pass_all": false,
    "repeats_completed": 23,
    "mismatch_count": 1,
    "first_mismatch": {
      "index": 22,
      "request_id": null,
      "text": "blue, green, red, yellow",
      "normalized": "blue, green, red, yellow",
      "token_ids": [
        11855,
        11,
        5983,
        11,
        2438,
        11,
        13358
      ],
      "finish_reason": "stop",
      "elapsed_s": 0.19228920798923355,
      "usage": {
        "prompt_tokens": 30,
        "total_tokens": 38,
        "completion_tokens": 8,
        "prompt_tokens_details": null
      },
      "pass": false
    }
  },
  "decision": {
    "accepted_by_requested_gates": false,
    "gate_failures": [
      "json_canary",
      "color_canary"
    ]
  }
}
