{
  "label": "tp4-mtp-k1-parity-fix-v2",
  "stamp": "20260620061440",
  "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-parity-fix-v2-p512o512-20260620061440.json",
    "json_canary": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-parity-fix-v2-json-repeat96-20260620061440.json",
    "color_canary": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-parity-fix-v2-color-repeat96-20260620061440.json",
    "quality_suite": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-parity-fix-v2-quality-suite-20260620061440.json",
    "log": "/home/steve/llm-optimizations/data/qwen36-ablation-tp4-mtp-k1-parity-fix-v2-20260620061440.log"
  },
  "artifact_exists": {
    "metrics": true,
    "json_canary": true,
    "color_canary": true,
    "quality_suite": false,
    "log": true
  },
  "env": {
    "MODEL_PATH": "/mnt/fast-ai/qwen36-quark-int8-fp8-mtp-hybrid",
    "SERVED_MODEL_NAME": "qwen36-35b-a3b-fp8",
    "SERVER_LAUNCHER": "scripts/launch-qwen36-quark-int8-accepted.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-parity-fix-v2",
    "TORCHINDUCTOR_CACHE_DIR": "/mnt/fast-ai/vllm-cache-exp/qwen36-ablation-tp4-mtp-k1-parity-fix-v2/torchinductor",
    "VLLM_CACHE_ROOT": "/mnt/fast-ai/vllm-cache-exp/qwen36-ablation-tp4-mtp-k1-parity-fix-v2/vllm",
    "BASE_URL": "http://127.0.0.1:18080",
    "PORT": "18080",
    "VLLM_XPU_GDN_NATIVE_FALLBACK": "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": "1",
    "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": "1",
    "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": "1",
    "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\":128}",
    "SPEC_CONFIG": null,
    "ABLATION_FAST_GRAPH_AUTOCONFIG": "0",
    "VLLM_EXTRA_ARGS": "--speculative-config {\"method\":\"mtp\",\"num_speculative_tokens\":1}",
    "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": 107.97654202272932,
      "median": 107.97654202272932,
      "min": 102.43259140781036,
      "max": 113.52049263764829
    },
    "tok_s_out_client_after_first_chunk_corrected": {
      "mean": 107.76565033909118,
      "median": 107.76565033909118,
      "min": 102.23252775271698,
      "max": 113.29877292546539
    },
    "tok_s_out_client_e2e": {
      "mean": 102.14490404563159,
      "median": 102.14490404563159,
      "min": 97.18693572438421,
      "max": 107.10287236687898
    },
    "tok_s_total_client": {
      "mean": 201.49678337126545,
      "median": 201.49678337126545,
      "min": 191.7164161750548,
      "max": 211.27715056747613
    },
    "ttft_ms_client": {
      "mean": 270.0202294990959,
      "median": 270.0202294990959,
      "min": 269.78866099852894,
      "max": 270.2517979996628
    },
    "ttft_ms_vllm_metrics": {
      "mean": 251.0848045349121,
      "median": 251.0848045349121,
      "min": 250.37050247192383,
      "max": 251.7991065979004
    },
    "tok_s_prefill_lower_bound_from_ttft": {
      "mean": 1844.3076179489658,
      "median": 1844.3076179489658,
      "min": 1842.7259455295884,
      "max": 1845.889290368343
    },
    "e2e_ms_vllm_metrics": {
      "mean": 5023.064613342285,
      "median": 5023.064613342285,
      "min": 4779.198169708252,
      "max": 5266.931056976318
    },
    "queue_ms_vllm_histogram": {
      "mean": 0.011001000530086458,
      "median": 0.011001000530086458,
      "min": 0.00902700048754923,
      "max": 0.012975000572623685
    },
    "prefill_ms_vllm_histogram": {
      "mean": 244.30397199921572,
      "median": 244.30397199921572,
      "min": 242.03861599926313,
      "max": 246.5693279991683
    },
    "decode_ms_vllm_histogram": {
      "mean": 4772.073827500208,
      "median": 4772.073827500208,
      "min": 4527.56539999973,
      "max": 5016.582255000685
    },
    "decode_ms_per_generation_token_vllm_histogram": {
      "mean": 9.320456694336343,
      "median": 9.320456694336343,
      "min": 8.842901171874473,
      "max": 9.798012216798213
    },
    "inference_ms_vllm_histogram": {
      "mean": 5016.377799499423,
      "median": 5016.377799499423,
      "min": 4774.134727998899,
      "max": 5258.620870999948
    },
    "time_per_output_token_ms_vllm_histogram": {
      "mean": 9.338696335616845,
      "median": 9.338696335616845,
      "min": 8.860206262230392,
      "max": 9.817186409003298
    },
    "inter_token_ms_vllm_histogram": {
      "mean": 16.65321474734497,
      "median": 16.65321474734497,
      "min": 16.584488644687656,
      "max": 16.721940850002284
    },
    "iteration_tokens_per_step_vllm_histogram": {
      "mean": 3.520806557218032,
      "median": 3.520806557218032,
      "min": 3.355481727574751,
      "max": 3.6861313868613137
    }
  },
  "speed": {
    "tok_s_out_corrected_mean": 107.76565033909118,
    "tok_s_out_after_first_chunk_mean": 107.97654202272932,
    "tok_s_out_e2e_mean": 102.14490404563159,
    "decode_ms_per_token_mean": 9.320456694336343,
    "ttft_ms_client_mean": 270.0202294990959
  },
  "json_canary": {
    "pass_all": false,
    "repeats_completed": 1,
    "mismatch_count": 1,
    "first_mismatch": {
      "index": 0,
      "request_id": null,
      "text": "{\"4404344.{\"44434040.{\"422}}{\"420}",
      "normalized": "{\"4404344.{\"44434040.{\"422}}{\"420}",
      "token_ids": [
        4754,
        19,
        19,
        15,
        19,
        18,
        19,
        19,
        13,
        4754,
        19,
        19,
        19,
        18,
        19,
        15,
        19,
        15,
        13,
        4754,
        19,
        17,
        17,
        3307,
        4754,
        19,
        17,
        15,
        92
      ],
      "finish_reason": "stop",
      "elapsed_s": 0.5884255820001272,
      "usage": {
        "prompt_tokens": 36,
        "total_tokens": 66,
        "completion_tokens": 30,
        "prompt_tokens_details": null
      },
      "json_parsed": null,
      "pass": false
    }
  },
  "color_canary": {
    "pass_all": false,
    "repeats_completed": 1,
    "mismatch_count": 1,
    "first_mismatch": {
      "index": 0,
      "request_id": null,
      "text": "blue, green, red, yellow,",
      "normalized": "blue, green, red, yellow,",
      "token_ids": [
        11855,
        11,
        5983,
        11,
        2438,
        11,
        13358,
        11
      ],
      "finish_reason": "stop",
      "elapsed_s": 0.23539511099988886,
      "usage": {
        "prompt_tokens": 30,
        "total_tokens": 39,
        "completion_tokens": 9,
        "prompt_tokens_details": null
      },
      "pass": false
    }
  },
  "decision": {
    "accepted_by_requested_gates": false,
    "gate_failures": [
      "json_canary",
      "color_canary"
    ]
  }
}
