[
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/awq-bf16-eager-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/awq-bf16-eager-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 13.603940153028816,
    "gpu_reserved_peak_mib": 22244.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.416,
        "decode_median_tok_s_estimate": 17.381,
        "e2e_median_tok_s": 10.538,
        "aggregate_median_tok_s": 10.535,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 4.396,
        "decode_median_tok_s_estimate": 17.134,
        "e2e_median_tok_s": 5.774,
        "aggregate_median_tok_s": 11.6,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/awq-bf16-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/awq-bf16-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 5.0573566580424085,
    "gpu_reserved_peak_mib": 22320.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 0.976,
        "decode_median_tok_s_estimate": 117.618,
        "e2e_median_tok_s": 29.214,
        "aggregate_median_tok_s": 29.199,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.935,
        "decode_median_tok_s_estimate": 110.116,
        "e2e_median_tok_s": 16.639,
        "aggregate_median_tok_s": 31.84,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/baseline-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/baseline-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 17.783527772990055,
    "gpu_reserved_peak_mib": 19885.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3922,
        "prompt_tokens_max": 3922,
        "ttft_median_seconds": 1.262,
        "decode_median_tok_s_estimate": 13.915,
        "e2e_median_tok_s": 10.623,
        "aggregate_median_tok_s": 10.621,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3922,
        "prompt_tokens_max": 3922,
        "ttft_median_seconds": 2.219,
        "decode_median_tok_s_estimate": 12.75,
        "e2e_median_tok_s": 8.596,
        "aggregate_median_tok_s": 13.97,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/baseline-recheck-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/baseline-recheck-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 15.627091745962389,
    "gpu_reserved_peak_mib": 19831.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.119,
        "decode_median_tok_s_estimate": 15.792,
        "e2e_median_tok_s": 12.039,
        "aggregate_median_tok_s": 12.037,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.146,
        "decode_median_tok_s_estimate": 14.974,
        "e2e_median_tok_s": 9.813,
        "aggregate_median_tok_s": 15.948,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/dflash-16-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/dflash-16-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 300,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 4.704084233031608,
    "gpu_reserved_peak_mib": 19436.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.008,
        "decode_median_tok_s_estimate": 404.585,
        "e2e_median_tok_s": 34.546,
        "aggregate_median_tok_s": 34.525,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.227,
        "decode_median_tok_s_estimate": 385.182,
        "e2e_median_tok_s": 16.4,
        "aggregate_median_tok_s": 33.185,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/dflash-4-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/dflash-4-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 300,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.303990040090866,
    "gpu_reserved_peak_mib": 17032.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.2,
        "decode_median_tok_s_estimate": 161.713,
        "e2e_median_tok_s": 26.596,
        "aggregate_median_tok_s": 26.582,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.243,
        "decode_median_tok_s_estimate": 162.153,
        "e2e_median_tok_s": 15.483,
        "aggregate_median_tok_s": 30.831,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/dflash-8-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/dflash-8-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 300,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 4.710330746951513,
    "gpu_reserved_peak_mib": 17822.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.01,
        "decode_median_tok_s_estimate": 264.776,
        "e2e_median_tok_s": 33.055,
        "aggregate_median_tok_s": 33.037,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.102,
        "decode_median_tok_s_estimate": 270.225,
        "e2e_median_tok_s": 17.052,
        "aggregate_median_tok_s": 33.69,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/flashinfer-prefill-32k-warmkernel/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/flashinfer-prefill-32k-warmkernel",
      "lengths": "32768",
      "concurrency": "1",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 11.319138183956966,
    "gpu_reserved_peak_mib": 24080.0,
    "rows": [
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 1,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 1,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 10.146,
        "decode_median_tok_s_estimate": 69.25,
        "e2e_median_tok_s": 5.195,
        "aggregate_median_tok_s": 5.194,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/flashinfer-prefill-compare/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/flashinfer-prefill-compare",
      "lengths": "32768,131072",
      "concurrency": "1",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 84.76467836194206,
    "gpu_reserved_peak_mib": 24108.0,
    "rows": [
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 1,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 1,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 20.914,
        "decode_median_tok_s_estimate": 69.847,
        "e2e_median_tok_s": 1.771,
        "aggregate_median_tok_s": 1.771,
        "errors": []
      },
      {
        "context_budget": 131072,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 1,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 1,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 60.582,
        "decode_median_tok_s_estimate": 42.875,
        "e2e_median_tok_s": 0.618,
        "aggregate_median_tok_s": 0.618,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/full-int8-bf16-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/full-int8-bf16-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.979468196048401,
    "gpu_reserved_peak_mib": 13652.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.06,
        "decode_median_tok_s_estimate": 85.518,
        "e2e_median_tok_s": 32.982,
        "aggregate_median_tok_s": 32.97,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.107,
        "decode_median_tok_s_estimate": 84.615,
        "e2e_median_tok_s": 14.851,
        "aggregate_median_tok_s": 32.61,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/full-int8-kve5m2-ascending/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/full-int8-kve5m2-ascending",
      "lengths": "4096,32768,131072,262144",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 1933.7073087169556,
    "gpu_reserved_peak_mib": 23824.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.398,
        "decode_median_tok_s_estimate": 85.563,
        "e2e_median_tok_s": 20.603,
        "aggregate_median_tok_s": 20.596,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 0.22,
        "decode_median_tok_s_estimate": 83.888,
        "e2e_median_tok_s": 56.087,
        "aggregate_median_tok_s": 167.775,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 12.96,
        "decode_median_tok_s_estimate": 69.222,
        "e2e_median_tok_s": 4.134,
        "aggregate_median_tok_s": 4.134,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 26.104,
        "decode_median_tok_s_estimate": 3.969,
        "e2e_median_tok_s": 1.417,
        "aggregate_median_tok_s": 3.281,
        "errors": []
      },
      {
        "context_budget": 131072,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 106.127,
        "decode_median_tok_s_estimate": 42.88,
        "e2e_median_tok_s": 0.355,
        "aggregate_median_tok_s": 0.355,
        "errors": []
      },
      {
        "context_budget": 131072,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 212.426,
        "decode_median_tok_s_estimate": 41.379,
        "e2e_median_tok_s": 0.178,
        "aggregate_median_tok_s": 0.356,
        "errors": []
      },
      {
        "context_budget": 262144,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 261874,
        "prompt_tokens_max": 261874,
        "ttft_median_seconds": 355.781,
        "decode_median_tok_s_estimate": 28.104,
        "e2e_median_tok_s": 0.145,
        "aggregate_median_tok_s": 0.145,
        "errors": []
      },
      {
        "context_budget": 262144,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 261874,
        "prompt_tokens_max": 261874,
        "ttft_median_seconds": 711.399,
        "decode_median_tok_s_estimate": 0.143,
        "e2e_median_tok_s": 0.049,
        "aggregate_median_tok_s": 0.12,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/full-int8-kve5m2-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/full-int8-kve5m2-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 8.326828385936096,
    "gpu_reserved_peak_mib": 23808.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.544,
        "decode_median_tok_s_estimate": 86.068,
        "e2e_median_tok_s": 19.124,
        "aggregate_median_tok_s": 19.117,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 4.133,
        "decode_median_tok_s_estimate": 84.218,
        "e2e_median_tok_s": 8.282,
        "aggregate_median_tok_s": 18.516,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/mlp-int8-bf16-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/mlp-int8-bf16-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.947340433020145,
    "gpu_reserved_peak_mib": 16542.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.013,
        "decode_median_tok_s_estimate": 67.292,
        "e2e_median_tok_s": 24.059,
        "aggregate_median_tok_s": 24.048,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.012,
        "decode_median_tok_s_estimate": 66.301,
        "e2e_median_tok_s": 14.683,
        "aggregate_median_tok_s": 31.733,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/mtp-1-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/mtp-1-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 300,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.49032649199944,
    "gpu_reserved_peak_mib": 12906.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.064,
        "decode_median_tok_s_estimate": 113.151,
        "e2e_median_tok_s": 26.937,
        "aggregate_median_tok_s": 26.912,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.178,
        "decode_median_tok_s_estimate": 112.934,
        "e2e_median_tok_s": 15.058,
        "aggregate_median_tok_s": 29.241,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/mtp-2-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/mtp-2-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 300,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.216005841968581,
    "gpu_reserved_peak_mib": 13118.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.051,
        "decode_median_tok_s_estimate": 132.102,
        "e2e_median_tok_s": 28.536,
        "aggregate_median_tok_s": 28.518,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.046,
        "decode_median_tok_s_estimate": 108.327,
        "e2e_median_tok_s": 15.909,
        "aggregate_median_tok_s": 30.788,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/mtp-4-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/mtp-4-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 300,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 4.9528945399215445,
    "gpu_reserved_peak_mib": 13502.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.018,
        "decode_median_tok_s_estimate": 164.497,
        "e2e_median_tok_s": 30.562,
        "aggregate_median_tok_s": 30.538,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.055,
        "decode_median_tok_s_estimate": 164.918,
        "e2e_median_tok_s": 16.814,
        "aggregate_median_tok_s": 32.404,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/text-bf16-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/text-bf16-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 7.097114444011822,
    "gpu_reserved_peak_mib": 22342.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 0.924,
        "decode_median_tok_s_estimate": 47.106,
        "e2e_median_tok_s": 26.686,
        "aggregate_median_tok_s": 26.677,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.823,
        "decode_median_tok_s_estimate": 47.114,
        "e2e_median_tok_s": 14.474,
        "aggregate_median_tok_s": 31.773,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/text-fp8-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/text-fp8-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 6.544370444025844,
    "gpu_reserved_peak_mib": 22294.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.031,
        "decode_median_tok_s_estimate": 77.633,
        "e2e_median_tok_s": 32.246,
        "aggregate_median_tok_s": 32.235,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.064,
        "decode_median_tok_s_estimate": 76.328,
        "e2e_median_tok_s": 20.254,
        "aggregate_median_tok_s": 37.236,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/sglang/results/text-fp8-mlp-graphs-4k/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8130/v1",
      "model": "ornith-sglang",
      "out": "/home/lab/bench/ornith/modular/sglang/results/text-fp8-mlp-graphs-4k",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-f4a988cb-c310-b618-5a9c-4462d4d33add",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 6.758954374003224,
    "gpu_reserved_peak_mib": 22314.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 0.999,
        "decode_median_tok_s_estimate": 66.235,
        "e2e_median_tok_s": 30.621,
        "aggregate_median_tok_s": 30.609,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.979,
        "decode_median_tok_s_estimate": 65.221,
        "e2e_median_tok_s": 19.957,
        "aggregate_median_tok_s": 36.249,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/shared/llama-baseline/results.jsonl",
    "execution_ended": false,
    "completed_grid": null,
    "reported_all_pass": null,
    "arguments": {},
    "task": "retrieval",
    "wall_seconds": null,
    "gpu_reserved_peak_mib": null,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.769,
        "decode_median_tok_s_estimate": 80.202,
        "e2e_median_tok_s": 16.886,
        "aggregate_median_tok_s": 16.881,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 15,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 4.074,
        "decode_median_tok_s_estimate": 31.524,
        "e2e_median_tok_s": 9.533,
        "aggregate_median_tok_s": 25.247,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 12.814,
        "decode_median_tok_s_estimate": 78.073,
        "e2e_median_tok_s": 2.859,
        "aggregate_median_tok_s": 2.859,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 15,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 28.117,
        "decode_median_tok_s_estimate": 3.686,
        "e2e_median_tok_s": 0.993,
        "aggregate_median_tok_s": 2.977,
        "errors": []
      },
      {
        "context_budget": 131072,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 68.354,
        "decode_median_tok_s_estimate": 50.659,
        "e2e_median_tok_s": 0.55,
        "aggregate_median_tok_s": 0.55,
        "errors": []
      },
      {
        "context_budget": 131072,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 15,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 147.236,
        "decode_median_tok_s_estimate": 0.542,
        "e2e_median_tok_s": 0.182,
        "aggregate_median_tok_s": 0.527,
        "errors": []
      },
      {
        "context_budget": 262144,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 261874,
        "prompt_tokens_max": 261874,
        "ttft_median_seconds": 194.453,
        "decode_median_tok_s_estimate": 29.695,
        "e2e_median_tok_s": 0.194,
        "aggregate_median_tok_s": 0.194,
        "errors": []
      },
      {
        "context_budget": 262144,
        "http_concurrency": 3,
        "batches": 2,
        "requests": 6,
        "http_success": 6,
        "output_nonempty": 6,
        "retrieval_evaluated": 6,
        "retrieval_pass": 6,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 261874,
        "prompt_tokens_max": 261874,
        "ttft_median_seconds": 392.816,
        "decode_median_tok_s_estimate": 0.433,
        "e2e_median_tok_s": 0.09,
        "aggregate_median_tok_s": 0.208,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/awq-bf16-common/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/awq-bf16-common/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.371939870994538,
    "gpu_reserved_peak_mib": 22114.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.071,
        "decode_median_tok_s_estimate": 97.977,
        "e2e_median_tok_s": 26.021,
        "aggregate_median_tok_s": 26.009,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.55,
        "decode_median_tok_s_estimate": 32.196,
        "e2e_median_tok_s": 10.236,
        "aggregate_median_tok_s": 30.514,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/awq-bf16-eager/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/awq-bf16-eager/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 9.879789215978235,
    "gpu_reserved_peak_mib": 21976.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.076,
        "decode_median_tok_s_estimate": 14.441,
        "e2e_median_tok_s": 10.256,
        "aggregate_median_tok_s": 10.254,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.521,
        "decode_median_tok_s_estimate": 11.305,
        "e2e_median_tok_s": 6.477,
        "aggregate_median_tok_s": 18.973,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/awq-bf16-on/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/awq-bf16-on/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 5.304282199009322,
    "gpu_reserved_peak_mib": 22114.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.069,
        "decode_median_tok_s_estimate": 98.434,
        "e2e_median_tok_s": 26.096,
        "aggregate_median_tok_s": 26.083,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.532,
        "decode_median_tok_s_estimate": 33.176,
        "e2e_median_tok_s": 10.385,
        "aggregate_median_tok_s": 30.958,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/awq-kv4-eager/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/awq-kv4-eager/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 10.798900182009675,
    "gpu_reserved_peak_mib": 22122.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.163,
        "decode_median_tok_s_estimate": 12.897,
        "e2e_median_tok_s": 9.23,
        "aggregate_median_tok_s": 9.228,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.7,
        "decode_median_tok_s_estimate": 10.393,
        "e2e_median_tok_s": 5.989,
        "aggregate_median_tok_s": 17.534,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/awq-kv4-on/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/awq-kv4-on/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400
    },
    "task": "retrieval",
    "wall_seconds": 5.646473796921782,
    "gpu_reserved_peak_mib": 22116.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.114,
        "decode_median_tok_s_estimate": 93.452,
        "e2e_median_tok_s": 24.962,
        "aggregate_median_tok_s": 24.953,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.74,
        "decode_median_tok_s_estimate": 31.533,
        "e2e_median_tok_s": 9.678,
        "aggregate_median_tok_s": 28.856,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/awq-kv8-on/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/awq-kv8-on/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 5.531962653971277,
    "gpu_reserved_peak_mib": 22118.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.098,
        "decode_median_tok_s_estimate": 96.418,
        "e2e_median_tok_s": 25.439,
        "aggregate_median_tok_s": 25.427,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.632,
        "decode_median_tok_s_estimate": 30.923,
        "e2e_median_tok_s": 9.893,
        "aggregate_median_tok_s": 29.498,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-bf16/screening32k/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-bf16/screening32k",
      "lengths": "32768",
      "concurrency": "3",
      "repeats": 5,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 217.23018601804506,
    "gpu_reserved_peak_mib": 22078.0,
    "rows": [
      {
        "context_budget": 32768,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 15,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 40.084,
        "decode_median_tok_s_estimate": 60.112,
        "e2e_median_tok_s": 0.94,
        "aggregate_median_tok_s": 3.26,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-bf16-128k/screening128k-v2/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-bf16-128k/screening128k-v2",
      "lengths": "131072",
      "concurrency": "3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 325.9817205440486,
    "gpu_reserved_peak_mib": 23728.0,
    "rows": [
      {
        "context_budget": 131072,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 3,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 319.39,
        "decode_median_tok_s_estimate": 21.276,
        "e2e_median_tok_s": 0.118,
        "aggregate_median_tok_s": 0.355,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-bf16-250w/throughput-five/results.jsonl",
    "execution_ended": false,
    "completed_grid": null,
    "reported_all_pass": null,
    "arguments": {},
    "task": "retrieval",
    "wall_seconds": null,
    "gpu_reserved_peak_mib": null,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 5,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 5,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 1.444,
        "decode_median_tok_s_estimate": 62.717,
        "e2e_median_tok_s": 53.347,
        "aggregate_median_tok_s": 53.342,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 3,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 3,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 4.299,
        "decode_median_tok_s_estimate": 58.324,
        "e2e_median_tok_s": 39.203,
        "aggregate_median_tok_s": 117.593,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-dflash-b16/short-throughput/results.jsonl",
    "execution_ended": true,
    "completed_grid": false,
    "reported_all_pass": false,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-dflash-b16/short-throughput",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 512,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 85009,
      "task": "throughput"
    },
    "task": "throughput",
    "wall_seconds": 43.56859595095739,
    "gpu_reserved_peak_mib": 22570.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 1,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 1,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 1.115,
        "decode_median_tok_s_estimate": 89.702,
        "e2e_median_tok_s": 75.167,
        "aggregate_median_tok_s": 75.16,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 0,
        "output_nonempty": 3,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 3,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 3,
        "prompt_tokens_min": null,
        "prompt_tokens_max": null,
        "ttft_median_seconds": 3.521,
        "decode_median_tok_s_estimate": null,
        "e2e_median_tok_s": null,
        "aggregate_median_tok_s": 0.0,
        "errors": [
          "missing SSE [DONE]; missing finish_reason; missing positive completion token usage"
        ]
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-dflash-b4/short-throughput/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-dflash-b4/short-throughput",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 512,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 85009,
      "task": "throughput"
    },
    "task": "throughput",
    "wall_seconds": 35.46480949292891,
    "gpu_reserved_peak_mib": 22488.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 1,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 1,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 1.112,
        "decode_median_tok_s_estimate": 86.384,
        "e2e_median_tok_s": 72.859,
        "aggregate_median_tok_s": 72.849,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 3,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 3,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 3.658,
        "decode_median_tok_s_estimate": 20.905,
        "e2e_median_tok_s": 18.167,
        "aggregate_median_tok_s": 54.332,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-dflash-b8/short-throughput/results.jsonl",
    "execution_ended": true,
    "completed_grid": true,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/full-int8-dflash-b8/short-throughput",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 512,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 85009,
      "task": "throughput"
    },
    "task": "throughput",
    "wall_seconds": 38.82824087992776,
    "gpu_reserved_peak_mib": 22524.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 1,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 1,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 1.109,
        "decode_median_tok_s_estimate": 77.196,
        "e2e_median_tok_s": 66.25,
        "aggregate_median_tok_s": 66.244,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 0,
        "retrieval_pass": 0,
        "cross_request_marker_evaluated": 3,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 3,
        "prompt_tokens_min": 3572,
        "prompt_tokens_max": 3572,
        "ttft_median_seconds": 3.32,
        "decode_median_tok_s_estimate": 18.673,
        "e2e_median_tok_s": 16.685,
        "aggregate_median_tok_s": 49.666,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv4-262k-c3-v2/full-grid/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": false,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv4-262k-c3-v2/full-grid",
      "lengths": "4096,32768,131072,262144",
      "concurrency": "1,3",
      "repeats": 5,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 308.06489295500796,
    "gpu_reserved_peak_mib": 21992.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.652,
        "decode_median_tok_s_estimate": 67.229,
        "e2e_median_tok_s": 22.789,
        "aggregate_median_tok_s": 22.783,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 15,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 4.445,
        "decode_median_tok_s_estimate": 38.374,
        "e2e_median_tok_s": 9.597,
        "aggregate_median_tok_s": 28.781,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 4,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 18.569,
        "decode_median_tok_s_estimate": 57.672,
        "e2e_median_tok_s": 5.978,
        "aggregate_median_tok_s": 5.978,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 3,
        "batches": 3,
        "requests": 9,
        "http_success": 9,
        "output_nonempty": 9,
        "retrieval_evaluated": 9,
        "retrieval_pass": 6,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 46.675,
        "decode_median_tok_s_estimate": 36.976,
        "e2e_median_tok_s": 2.779,
        "aggregate_median_tok_s": 6.854,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv8-262k-c3/full-grid/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": false,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv8-262k-c3/full-grid",
      "lengths": "4096,32768,131072,262144",
      "concurrency": "1,3",
      "repeats": 5,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 576.1292485559825,
    "gpu_reserved_peak_mib": 22004.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.496,
        "decode_median_tok_s_estimate": 74.023,
        "e2e_median_tok_s": 25.145,
        "aggregate_median_tok_s": 25.136,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 15,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 4.21,
        "decode_median_tok_s_estimate": 39.912,
        "e2e_median_tok_s": 10.131,
        "aggregate_median_tok_s": 30.314,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 5,
        "requests": 5,
        "http_success": 5,
        "output_nonempty": 5,
        "retrieval_evaluated": 5,
        "retrieval_pass": 5,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 18.957,
        "decode_median_tok_s_estimate": 59.559,
        "e2e_median_tok_s": 3.51,
        "aggregate_median_tok_s": 3.51,
        "errors": []
      },
      {
        "context_budget": 32768,
        "http_concurrency": 3,
        "batches": 5,
        "requests": 15,
        "http_success": 15,
        "output_nonempty": 15,
        "retrieval_evaluated": 15,
        "retrieval_pass": 8,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 48.351,
        "decode_median_tok_s_estimate": 33.793,
        "e2e_median_tok_s": 4.511,
        "aggregate_median_tok_s": 10.165,
        "errors": []
      },
      {
        "context_budget": 131072,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 130804,
        "prompt_tokens_max": 130804,
        "ttft_median_seconds": 149.07,
        "decode_median_tok_s_estimate": 41.181,
        "e2e_median_tok_s": 0.253,
        "aggregate_median_tok_s": 0.253,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv8-262k-c3/same-seed/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv8-262k-c3/same-seed",
      "lengths": "32768",
      "concurrency": "1",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4600,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 20.67294502304867,
    "gpu_reserved_peak_mib": 21848.0,
    "rows": [
      {
        "context_budget": 32768,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 19.437,
        "decode_median_tok_s_estimate": 57.182,
        "e2e_median_tok_s": 2.556,
        "aggregate_median_tok_s": 2.555,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv8-ssm32-32k/screening32k/results.jsonl",
    "execution_ended": true,
    "completed_grid": false,
    "reported_all_pass": false,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/nf4-kv8-ssm32-32k/screening32k",
      "lengths": "32768",
      "concurrency": "3",
      "repeats": 5,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 115.4230415330967,
    "gpu_reserved_peak_mib": 22004.0,
    "rows": [
      {
        "context_budget": 32768,
        "http_concurrency": 3,
        "batches": 2,
        "requests": 6,
        "http_success": 6,
        "output_nonempty": 6,
        "retrieval_evaluated": 6,
        "retrieval_pass": 4,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 32500,
        "prompt_tokens_max": 32500,
        "ttft_median_seconds": 50.684,
        "decode_median_tok_s_estimate": 39.341,
        "e2e_median_tok_s": 3.39,
        "aggregate_median_tok_s": 8.846,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/original-bf16-kv4-on/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/original-bf16-kv4-on/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 7.380632409011014,
    "gpu_reserved_peak_mib": 23302.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.096,
        "decode_median_tok_s_estimate": 42.52,
        "e2e_median_tok_s": 23.374,
        "aggregate_median_tok_s": 23.367,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 2.664,
        "decode_median_tok_s_estimate": 27.444,
        "e2e_median_tok_s": 12.048,
        "aggregate_median_tok_s": 35.775,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  },
  {
    "source": "/home/lab/bench/ornith/modular/vllm/runs/original-nf4-kv4-on/bench/results.jsonl",
    "execution_ended": true,
    "completed_grid": null,
    "reported_all_pass": true,
    "arguments": {
      "base": "http://127.0.0.1:8140/v1",
      "model": "ornith-vllm",
      "out": "/home/lab/bench/ornith/modular/vllm/runs/original-nf4-kv4-on/bench",
      "lengths": "4096",
      "concurrency": "1,3",
      "repeats": 1,
      "max_output": 256,
      "timeout": 1800,
      "tokenizer": "/home/lab/models/Ornith-1.5-9B-original",
      "gpu": "GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f",
      "duration": 0,
      "seed": 4400,
      "task": "retrieval"
    },
    "task": "retrieval",
    "wall_seconds": 7.348940714029595,
    "gpu_reserved_peak_mib": 22102.0,
    "rows": [
      {
        "context_budget": 4096,
        "http_concurrency": 1,
        "batches": 1,
        "requests": 1,
        "http_success": 1,
        "output_nonempty": 1,
        "retrieval_evaluated": 1,
        "retrieval_pass": 1,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 1.221,
        "decode_median_tok_s_estimate": 72.537,
        "e2e_median_tok_s": 28.369,
        "aggregate_median_tok_s": 28.36,
        "errors": []
      },
      {
        "context_budget": 4096,
        "http_concurrency": 3,
        "batches": 1,
        "requests": 3,
        "http_success": 3,
        "output_nonempty": 3,
        "retrieval_evaluated": 3,
        "retrieval_pass": 3,
        "cross_request_marker_evaluated": 0,
        "cross_request_marker_failures": 0,
        "sse_terminal_recorded": 0,
        "prompt_tokens_min": 3828,
        "prompt_tokens_max": 3828,
        "ttft_median_seconds": 3.032,
        "decode_median_tok_s_estimate": 26.831,
        "e2e_median_tok_s": 11.088,
        "aggregate_median_tok_s": 33.035,
        "errors": []
      }
    ],
    "caveats": [
      "HTTP concurrency does not itself prove simultaneous GPU residency.",
      "SSE-derived decode rate is an estimate; chunking can batch output tokens.",
      "GPU memory is sampled total including reserved cache, not weight size."
    ]
  }
]
