{
  "date": "2026-07-29",
  "host": "DGX Spark GB10",
  "mode": "dflash",
  "decision": "DO_NOT_PROMOTE",
  "reason": "gen128 and prefill both slower than pinned Q4 baseline",
  "baseline_pin": {
    "gen128_tok_s": 21.47,
    "agent_smoke": "40/40",
    "hermes_v2": "27/27",
    "engine": "04b2b72",
    "weight": "laguna-s-2.1-Q4_K_M.gguf",
    "sha256": "a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4"
  },
  "dflash_measure": {
    "throughput": {
      "ts": 1785314162.0283267,
      "mode": "dflash",
      "spec": {
        "type": "draft-dflash",
        "n_max": 7,
        "block_size": 16
      },
      "base_url": "http://127.0.0.1:8000/v1",
      "model": "local-laguna",
      "rows": [
        {
          "label": "gen8_short",
          "prompt_tokens": 48,
          "completion_tokens": 3,
          "latency_s": 0.483,
          "tok_s": 6.212,
          "content_head": "OK"
        },
        {
          "label": "gen128",
          "prompt_tokens": 58,
          "completion_tokens": 128,
          "latency_s": 8.374,
          "tok_s": 15.286,
          "content_head": "Speculative decoding is a model inference optimization technique that accelerates text generation by using a fast, light"
        },
        {
          "label": "gen256",
          "prompt_tokens": 61,
          "completion_tokens": 256,
          "latency_s": 14.762,
          "tok_s": 17.342,
          "content_head": "## Block-Diffusion vs EAGLE-3: Practical Serving Comparison\n\n### **Block-Diffusion Draft Models**\n\n**What it is:** Uses "
        }
      ],
      "schema": "laguna.dflash_throughput/v1"
    },
    "server_bench": {
      "base_url": "http://127.0.0.1:8000/v1",
      "model": "local-laguna",
      "mode": "dflash",
      "rows": [
        {
          "mark": "2k",
          "latency_s": 2.253,
          "prompt_tokens": 836,
          "completion_tokens": 3,
          "completion_tok_s": 1.332,
          "content": "OK"
        },
        {
          "mark": "8k",
          "latency_s": 5.556,
          "prompt_tokens": 3236,
          "completion_tokens": 3,
          "completion_tok_s": 0.54,
          "content": "OK"
        }
      ],
      "ts": 1785314184.3951924
    },
    "metrics_snapshot": {
      "prompt_tokens_seconds": 458.721,
      "predicted_tokens_seconds": 17.288,
      "n_decode_total": 188,
      "tokens_predicted_total": 409
    },
    "props_notes": {
      "speculative_default": "none",
      "slot_speculative": true,
      "slot_speculative_types": "none,draft-dflash",
      "model_alias": "local-laguna",
      "n_ctx": 8192
    },
    "ps_args_head": "/home/hizrianraz/src/llama.cpp-laguna/build/bin/llama-server -m /home/hizrianraz/models/laguna-s-2.1/laguna-s-2.1-Q4_K_M.gguf -md /home/hizrianraz/models/laguna-s-2.1/laguna-s-2.1-DFlash-BF16.gguf --spec-type draft-dflash --spec-draft-n-max 7 --host 127.0.0.1 --port 8000 --ctx-size 8192 -ngl 99 -fa on --jinja --alias local-laguna --metrics"
  },
  "comparison": {
    "gen128_dflash_tok_s": 15.286,
    "gen128_baseline_tok_s": 21.47,
    "gen128_delta_tok_s": -6.184,
    "gen128_vs_baseline_pct": -28.8,
    "prefill_2k_dflash_s": 2.253,
    "prefill_8k_dflash_s": 5.556,
    "prefill_2k_baseline_s": 1.597,
    "prefill_8k_baseline_s": 4.78
  },
  "verdict": "DFlash DO_NOT_PROMOTE: gen128 15.286 t/s vs baseline 21.47 (-28.8%); 2k 2.253s vs ~1.6s; 8k 5.556s vs ~4.78s.",
  "ts_unix": 1785314326.7548192
}