{
  "schema_version": 1,
  "name": "efficiency-h200-best-v1",
  "measured": "2026-10-01",
  "comparison_scope": "Each before/after pair uses one H200 and one fixed panel. Absolute latency and accuracy must not be compared across panels or against the A100 comparison.",
  "rows": [
    {
      "model": "JevAny-Qwen3.8-27B",
      "hardware": "1 x NVIDIA H200",
      "panel": {
        "suite": "JevBench public",
        "records": 231
      },
      "before": {
        "configuration": "BF16 merged LoRA + FLA + fused SDPA; CUDA graphs disabled",
        "median_ms": 113.54,
        "correct": 207,
        "total": 231
      },
      "after": {
        "configuration": "same matched path; CUDA graphs enabled",
        "median_ms": 30.53,
        "correct": 207,
        "total": 231
      },
      "median_speedup": 3.72,
      "argmax_flips": 0,
      "source": "docs/DEPLOYMENT.md native CUDA Graph measurement"
    },
    {
      "model": "JevAny-Muse-Glimmer-30B",
      "hardware": "1 x NVIDIA H200",
      "panel": {
        "suite": "Transfer balanced sample",
        "records": 44
      },
      "before": {
        "configuration": "unmerged LoRA + math SDPA; CUDA graphs disabled",
        "median_ms": 100.71,
        "correct": 38,
        "total": 44
      },
      "intermediate": {
        "configuration": "unmerged LoRA + fused SDPA; CUDA graphs disabled",
        "median_ms": 87.6799295656383,
        "median_speedup_from_before": 1.1485907560796185
      },
      "after": {
        "configuration": "unmerged LoRA + fused SDPA + CUDA graphs",
        "median_ms": 43.25,
        "correct": 38,
        "total": 44,
        "median_speedup_from_intermediate": 2.0273546618482214,
        "graph_calls": 96,
        "eager_calls": 0
      },
      "median_speedup": 2.33,
      "argmax_flips": 0,
      "source": "single-H200 matched audit",
      "code_revision": "f5110b8b87b63f230f26c9cecd5fd13968bc4463",
      "checkpoint": {
        "base_revision": "a4e59da52a7bc87ae7251dd5545c0dd437c44b68",
        "adapter_sha256": "87434122ffa2cc9438baf1f0d6375fec75f47f246d070eb50eb2a543859c4832",
        "head_sha256": "c19591f6293c52ab796f85e5db6e34fc08f5cfbec294614e16a76d2c476f69b1"
      },
      "panel_sha256": "0b6a56bd063bdf82fbecd5bb77c61f070f3c0168b6ce9f07b838147a3eafc511",
      "raw_comparison_sha256": "aa532515e591a5d518d66d82bbfd3c05f5b7db126bcf4f8fe2629dd844410218"
    }
  ]
}