{
  "kind": "gemma4_geglu_down_matmul_epilogue_current_negative",
  "date": "2026-07-01",
  "status": "closed_negative",
  "localmaxxing_submission": false,
  "reason_not_submitted": "Semantically valid strict fresh-response runs, but the candidate was much slower than paired controls.",
  "source_worktree": "/home/steve/src/llama.cpp-gemma-record-repro-c926",
  "source_identity": "llama.cpp c926ad098 dirty Gemma/B70 optimization stack",
  "runtime": "/home/steve/src/llama.cpp-gemma-record-repro-c926/build-sycl-b70-aot-bmg-g31-q8reorder-vdr2/bin/llama-server",
  "target_model": "/mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf/gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf",
  "draft_model": "/mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf/MTP/gemma-4-26B-A4B-it-Q4_0-MTP.gguf",
  "quality_and_validity": {
    "fixed_realistic_suite": true,
    "each_prompt_once_cold": true,
    "cached_tokens_all_zero": true,
    "canary_repeats": 32,
    "max_tokens": 128,
    "primary_metric": "median generated-token tok/s for tokens 1-100 after TTFT"
  },
  "control_env": {
    "FLASH_ATTN": "on",
    "CTX_SIZE": "32768",
    "GGML_SYCL_ENABLE_VMM": "1",
    "BATCH_SIZE": "1024",
    "UBATCH_SIZE": "1024",
    "REALISTIC_GATE": "1",
    "REALISTIC_METRIC_TOKENS": "100"
  },
  "candidate_env": {
    "LLAMA_GEMMA4_MOE_GEGLU_DOWN_MATMUL_EPILOGUE": "1"
  },
  "rows": [
    {
      "label": "gemma4-q8-gpu0-gegludownmat-control-strict128-20260701T061422Z",
      "gpu": 0,
      "candidate": false,
      "median_1_100_tok_s": 119.05514562919211,
      "p10_1_100_tok_s": 108.34566452639842,
      "mean_1_100_tok_s": 119.64069818966999,
      "full128_median_tok_s": 120.51443228522186,
      "ttft_median_ms": 179.57178800133988
    },
    {
      "label": "gemma4-q8-gpu2-gegludownmat-control-strict128-20260701T061422Z",
      "gpu": 2,
      "candidate": false,
      "median_1_100_tok_s": 116.97487828110022,
      "p10_1_100_tok_s": 105.90600389251878,
      "mean_1_100_tok_s": 116.5344675542113,
      "full128_median_tok_s": 113.9073890041566,
      "ttft_median_ms": 179.1530919726938
    },
    {
      "label": "gemma4-q8-gpu1-gegludownmat-on-strict128-20260701T061422Z",
      "gpu": 1,
      "candidate": true,
      "median_1_100_tok_s": 78.88908260591842,
      "p10_1_100_tok_s": 69.96527099701746,
      "mean_1_100_tok_s": 78.95909931116093,
      "full128_median_tok_s": 77.62427191559408,
      "ttft_median_ms": 194.15506394580007
    },
    {
      "label": "gemma4-q8-gpu3-gegludownmat-on-strict128-20260701T061422Z",
      "gpu": 3,
      "candidate": true,
      "median_1_100_tok_s": 76.56989021267371,
      "p10_1_100_tok_s": 68.24174503126903,
      "mean_1_100_tok_s": 76.66715824269423,
      "full128_median_tok_s": 74.88630808257571,
      "ttft_median_ms": 194.0963984816335
    }
  ],
  "decision": "Do not promote, submit, or run full512 for this flag. It is roughly 38-42 tok/s slower than paired controls on the primary median metric."
}