{
  "artifacts": [
    {
      "file": "model_execmeta.litertlm",
      "sha256": "7ed2d8f9cc41eff47a0f2e21a5908b51c859bfbdf0a86c6403997d9caa49329d",
      "size_mb": 1189.481
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "python hf-to-litertlm/lfm_work/convert_lfm25.py LiquidAI/LFM2.5-1.2B-Instruct out_wi8_composite_conv031 --keep-softmax-composite",
    "quantization": "int8 dynamic weights / fp32 activations (litert-torch recipe dynamic_wi8_afp32)",
    "tool": "litert-torch (via hf-to-litertlm lfm_work/convert_lfm25.py + litertlm-convert scripts/add_executor_metadata.py)",
    "tool_version": "0.9.3 (litert-converter 0.3.1; venv ~/venvs/ltconv040dev per RESULTS.md 2026-08-11)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-08-12",
          "decode_tokens_per_s": 262.55,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": null,
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "results block: prefill=4643.93 decode=262.55 tokens/s, init=1.3371 s"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 256.0,
            "init_s": 1.3371,
            "max_num_tokens": 1024.0,
            "prefill_tokens": 256.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 4643.93,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 58.9
        },
        "source": "data/device_runs/0.16.0/2026-08-12/lfm25-12b-wi8-composite__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "lfm2.5",
    "id": "lfm25-12b-wi8-composite",
    "license": "lfm-open-license-v1.0",
    "source_url": "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct",
    "task": "text-generation"
  },
  "pitfalls": [
    "This variant deliberately KEEPS the odml.softmax composites (--keep-softmax-composite) to verify litert-converter 0.3.1 lowers them: 0.3.0 left 216 markers, 0.3.1 leaves 0, and the op histogram is IDENTICAL to the 0.4.0.dev20260806 baseline. Marker counts alone are not evidence — dead metadata can remain; compare op histograms (RESULTS.md 手順 2-3).",
    "litert-torch 0.9.3 does not write the ExecutorMetadata section (unlike main); litert-lm >=0.15 requires it to bind the hybrid ShortConv/attention state buffers (22 here) — retrofitted with scripts/add_executor_metadata.py (RESULTS.md 手順 4, add_execmeta.log).",
    "litert-torch 0.9.3 pins litert-converter==0.3.*, so a fresh install today pulls the lowering 0.3.1 — but the same install on 2026-08-08 pulled 0.3.0 (no lowering). Check the converter version, not just litert-torch (RESULTS.md 環境).",
    "Mac GPU prefill numbers swing about +/-11% between sessions on op-identical graphs (4830 vs 4297 tok/s here); treat prefill deltas inside that band as measurement noise — decode and init are stable, and PASS/FAIL judgments are load-independent (RESULTS.md 数字の読み方)."
  ],
  "schema_version": "1.2"
}