{
  "artifacts": [
    {
      "file": "granite-4.0-h-1b_int8_fp32act.litertlm",
      "sha256": "c4c5d09334b974abd2f80893b1b46df036a8ff28b54e70d36309326f9f50fa06",
      "size_mb": 1605.158
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "PATH=$REPO/.venv-092/bin:$PATH PYTHONPATH=$HOME/code/litert-torch .venv-092/bin/python $HOME/code/hf-to-litertlm/granite_work/convert_granite4h.py ibm-granite/granite-4.0-h-1b <outdir>  # gpu_ship_20260813/build_1b_final.sh verbatim",
    "quantization": "int8 dynamic on linears + embedding; convs and the selective scan stay float; fp32 activations declared (GPU requirement — HF card update note)",
    "tool": "litert-torch, pinned base + granite hybrid patch, selective scan folded to batched matmuls (all tensors rank <= 4, no giant broadcast intermediates) — hf-to-litertlm granite_work/convert_granite4h.py via gpu_ship_20260813/build_1b_final.sh",
    "tool_version": "editable checkout 115a136 + granite_hybrid_litert_torch.patch (hf-to-litertlm granite_work/README.md; folded-scan revision of 2026-08-13)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-08-13",
          "decode_tokens_per_s": 4.34,
          "delegated_ops": null,
          "env": {
            "device": "pixel-8a",
            "machine_label": "pixel-8a-cl-pinned-thermally-loaded",
            "os_build": null,
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "Google Tensor G3",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "results block: prefill=5.14 decode=4.34 tokens/s"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 48.0,
            "init_s": 19.01252,
            "prefill_tokens": 40.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 5.14,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 8010.0
        },
        "source": "data/device_runs/0.16.0/2026-08-13/granite4h-1b-gpufix__pixel-8a.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-08-13",
          "decode_tokens_per_s": 8.81,
          "delegated_ops": 4488,
          "env": {
            "device": "pixel-8a",
            "machine_label": "pixel-8a-cl-pinned",
            "os_build": null,
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "Google Tensor G3",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "VERBOSE: Replacing 4346 out of 4346 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 4346 out of 4346 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "VERBOSE: Replacing 4310 out of 4310 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).",
            "VERBOSE: Replacing 4488 out of 4488 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).",
            "results block: prefill=23.9 decode=8.81 tokens/s"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 52.0,
            "init_s": 83.68662,
            "prefill_tokens": 40.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 23.9,
          "provenance": "measured",
          "runs": true,
          "total_ops": 4488,
          "ttft_ms": 1790.0
        },
        "source": "data/device_runs/0.16.0/2026-08-13/granite4h-1b-gpufix__pixel-8a.json"
      }
    ]
  },
  "model": {
    "family": "granite4h",
    "id": "granite4h-1b-gpufix",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/ibm-granite/granite-4.0-h-1b",
    "task": "text-generation"
  },
  "pitfalls": [
    "GPU execution requires fp32 activations (declared in this bundle) — the memory cost is stated on the HF card's honest notes; CPU decode is ~3.5x the previous file (11.4 -> 39.2 tok/s) from the folded scan alone (HF card update, measured).",
    "Full GPU delegation on Android OpenCL, one partition per subgraph, zero rejections (4628/4628, 2240/2240, 2271/2271) — litertlm-convert commit 2cd79bf; end-to-end GPU verified on macOS, iPhone 17 Pro (Metal) and Pixel 8a per the HF card update.",
    "First Mamba2-hybrid selective scan on a mobile GPU (HF card claim of record).",
    "Family BOS trap unchanged from the 08-04 lineage: the 350m sibling's greedy flips with BOS (LlmMetadata start_token); the 1b is BOS-robust (commit 5f3176d)."
  ],
  "schema_version": "1.2"
}