{
  "title": "MiniV DFlash Pilot Benchmark Results",
  "run_id": "dflash-matrix-20260708_153700",
  "date": "2026-07-08",
  "hardware_public": {
    "system": "Corsair AI Workstation 300 / MiniV",
    "apu": "AMD Ryzen AI MAX 385",
    "gpu_backend": "Radeon 8050S Graphics, RADV Vulkan / GFX1151",
    "memory": "64GB LPDDR5X UMA, 48GB iGPU allocation target"
  },
  "methodology_public": {
    "scope": "YAGNI pilot: one prompt, 8192 context, 256 generated tokens, Q8 DFlash drafters only.",
    "baseline": "Same isolated DFlash-capable llama.cpp Vulkan build with --spec-type none.",
    "dflash": "Same isolated build with --spec-type draft-dflash and Q8_0 drafter.",
    "build": "llama.cpp DFlash PR build e8f1015, Vulkan; not the production TurboQuant router build.",
    "important_caveat": "Responses produced 256 completion tokens but exposed empty content due thinking-template behavior; speed timings are valid, output quality is not evaluated in this pilot.",
    "metric_rule": "Decode throughput is reported from llama-server predicted_per_second and wall completion-tokens/elapsed separately; no combined prefill+decode TPS is used."
  },
  "decision": "Do not promote DFlash on MiniV yet. In this isolated Vulkan build, DFlash is slower than baseline across all tested compatible models despite real draft acceptance.",
  "rows": [
    {
      "run_id": "dflash-matrix-20260708_153700",
      "case_id": "qwen36-27b-q6__baseline",
      "model_id": "qwen36-27b-q6",
      "model_label": "Qwen3.6 / Qwopus3.6 27B Q6_K",
      "mode": "baseline",
      "ctx_tokens": 8192,
      "max_output_tokens": 256,
      "draft_n_max": 0,
      "status": "SPEED_VALID_EMPTY_CONTENT",
      "load_sec": 64.465,
      "prompt_tokens": 59,
      "completion_tokens": 256,
      "prefill_tps": 45.255,
      "decode_tps_llama_timing": 9.481,
      "decode_tps_wall": 9.021,
      "prompt_ms": 1303.732,
      "predicted_ms": 27002.187
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "case_id": "qwen36-27b-q6__dflash",
      "model_id": "qwen36-27b-q6",
      "model_label": "Qwen3.6 / Qwopus3.6 27B Q6_K",
      "mode": "dflash",
      "ctx_tokens": 8192,
      "max_output_tokens": 256,
      "draft_n_max": 8,
      "status": "SPEED_VALID_EMPTY_CONTENT",
      "load_sec": 70.16,
      "prompt_tokens": 59,
      "completion_tokens": 256,
      "prefill_tps": 41.807,
      "decode_tps_llama_timing": 8.337,
      "decode_tps_wall": 7.949,
      "prompt_ms": 1411.253,
      "predicted_ms": 30705.424,
      "draft_acceptance_rate": 0.30782,
      "draft_tokens_accepted": 181,
      "draft_tokens_generated": 588,
      "mean_acceptance_length": 3.45
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "case_id": "qwen36-35b-a3b-q6__baseline",
      "model_id": "qwen36-35b-a3b-q6",
      "model_label": "Qwen3.6 35B-A3B Q6_K_XL",
      "mode": "baseline",
      "ctx_tokens": 8192,
      "max_output_tokens": 256,
      "draft_n_max": 0,
      "status": "SPEED_VALID_EMPTY_CONTENT",
      "load_sec": 91.402,
      "prompt_tokens": 59,
      "completion_tokens": 256,
      "prefill_tps": 77.259,
      "decode_tps_llama_timing": 51.072,
      "decode_tps_wall": 43.802,
      "prompt_ms": 763.669,
      "predicted_ms": 5012.552
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "case_id": "qwen36-35b-a3b-q6__dflash",
      "model_id": "qwen36-35b-a3b-q6",
      "model_label": "Qwen3.6 35B-A3B Q6_K_XL",
      "mode": "dflash",
      "ctx_tokens": 8192,
      "max_output_tokens": 256,
      "draft_n_max": 8,
      "status": "SPEED_VALID_EMPTY_CONTENT",
      "load_sec": 94.191,
      "prompt_tokens": 59,
      "completion_tokens": 256,
      "prefill_tps": 83.797,
      "decode_tps_llama_timing": 33.717,
      "decode_tps_wall": 30.603,
      "prompt_ms": 704.081,
      "predicted_ms": 7592.696,
      "draft_acceptance_rate": 0.37402,
      "draft_tokens_accepted": 190,
      "draft_tokens_generated": 508,
      "mean_acceptance_length": 3.92
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "case_id": "gemma4-26b-a4b-q6__baseline",
      "model_id": "gemma4-26b-a4b-q6",
      "model_label": "Gemma4 26B-A4B-it Q6_K_XL",
      "mode": "baseline",
      "ctx_tokens": 8192,
      "max_output_tokens": 256,
      "draft_n_max": 0,
      "status": "SPEED_VALID_EMPTY_CONTENT",
      "load_sec": 64.135,
      "prompt_tokens": 67,
      "completion_tokens": 256,
      "prefill_tps": 176.638,
      "decode_tps_llama_timing": 44.831,
      "decode_tps_wall": 41.136,
      "prompt_ms": 379.307,
      "predicted_ms": 5710.298
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "case_id": "gemma4-26b-a4b-q6__dflash",
      "model_id": "gemma4-26b-a4b-q6",
      "model_label": "Gemma4 26B-A4B-it Q6_K_XL",
      "mode": "dflash",
      "ctx_tokens": 8192,
      "max_output_tokens": 256,
      "draft_n_max": 8,
      "status": "SPEED_VALID_EMPTY_CONTENT",
      "load_sec": 65.16,
      "prompt_tokens": 67,
      "completion_tokens": 256,
      "prefill_tps": 159.646,
      "decode_tps_llama_timing": 31.1,
      "decode_tps_wall": 29.207,
      "prompt_ms": 419.678,
      "predicted_ms": 8231.419,
      "draft_acceptance_rate": 0.34259,
      "draft_tokens_accepted": 185,
      "draft_tokens_generated": 540,
      "mean_acceptance_length": 3.64
    }
  ],
  "comparisons": [
    {
      "run_id": "dflash-matrix-20260708_153700",
      "model_id": "qwen36-27b-q6",
      "model_label": "Qwen3.6 / Qwopus3.6 27B Q6_K",
      "baseline_status": "SPEED_VALID_EMPTY_CONTENT",
      "dflash_status": "SPEED_VALID_EMPTY_CONTENT",
      "baseline_decode_tps_wall": 9.021,
      "dflash_decode_tps_wall": 7.949,
      "wall_speedup_ratio": 0.881,
      "wall_delta_pct": -11.9,
      "baseline_decode_tps_llama": 9.481,
      "dflash_decode_tps_llama": 8.337,
      "llama_speedup_ratio": 0.879,
      "llama_delta_pct": -12.1,
      "draft_acceptance_rate": 0.30782,
      "draft_tokens_accepted": 181,
      "draft_tokens_generated": 588,
      "mean_acceptance_length": 3.45
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "model_id": "qwen36-35b-a3b-q6",
      "model_label": "Qwen3.6 35B-A3B Q6_K_XL",
      "baseline_status": "SPEED_VALID_EMPTY_CONTENT",
      "dflash_status": "SPEED_VALID_EMPTY_CONTENT",
      "baseline_decode_tps_wall": 43.802,
      "dflash_decode_tps_wall": 30.603,
      "wall_speedup_ratio": 0.699,
      "wall_delta_pct": -30.1,
      "baseline_decode_tps_llama": 51.072,
      "dflash_decode_tps_llama": 33.717,
      "llama_speedup_ratio": 0.66,
      "llama_delta_pct": -34.0,
      "draft_acceptance_rate": 0.37402,
      "draft_tokens_accepted": 190,
      "draft_tokens_generated": 508,
      "mean_acceptance_length": 3.92
    },
    {
      "run_id": "dflash-matrix-20260708_153700",
      "model_id": "gemma4-26b-a4b-q6",
      "model_label": "Gemma4 26B-A4B-it Q6_K_XL",
      "baseline_status": "SPEED_VALID_EMPTY_CONTENT",
      "dflash_status": "SPEED_VALID_EMPTY_CONTENT",
      "baseline_decode_tps_wall": 41.136,
      "dflash_decode_tps_wall": 29.207,
      "wall_speedup_ratio": 0.71,
      "wall_delta_pct": -29.0,
      "baseline_decode_tps_llama": 44.831,
      "dflash_decode_tps_llama": 31.1,
      "llama_speedup_ratio": 0.694,
      "llama_delta_pct": -30.6,
      "draft_acceptance_rate": 0.34259,
      "draft_tokens_accepted": 185,
      "draft_tokens_generated": 540,
      "mean_acceptance_length": 3.64
    }
  ]
}