{
  "artifacts": [
    {
      "file": "Shieldstral-1.0-3B-vision_int4_gpu.litertlm",
      "sha256": "d4b1a34690ed991892bfa8c8dcbafc78a681cbcb9b18e6c8b4544fccdc748339",
      "size_mb": 2654.158
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "shieldstral_work/vision_gpu_20260827/build_decoder_gpu.sh -> build_shieldstral_bundle.py",
    "quantization": "int4 decoder + int8 vision adapter/tower (per bundle inputs in s5_gate/ARTIFACTS.md)",
    "tool": "litert-torch export (decoder re-export, composite-free) + bundle rebuild (build_shieldstral_bundle.py)",
    "tool_version": "decoder re-export on ~/venvs/ltconv040dev, py3.10.13: litert-torch 0.9.3 / litert-converter 0.3.1 / ai-edge-quantizer 0.8.0 — the converter is the lever, not the torch bump (0.3.1 lowers odml.softmax natively where 0.3.0 cannot). Bundle repack: TODO(owner) — build_shieldstral_bundle.py imports only litert_lm_builder, but the interpreter that ran it is unrecorded, so no builder version is establishable."
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-09-05",
          "decode_tokens_per_s": 1.33,
          "delegated_ops": 1433,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-cold-cache-cooled",
            "os_build": "Android 16",
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "VERBOSE: Replacing 1076 out of 1187 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 153 partitions for subgraph 0 (prefill_2048).",
            "VERBOSE: Replacing 1076 out of 1187 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 153 partitions for subgraph 1 (prefill_1024).",
            "VERBOSE: Replacing 1076 out of 1187 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 153 partitions for subgraph 2 (prefill_512).",
            "VERBOSE: Replacing 1076 out of 1187 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 153 partitions for subgraph 3 (prefill_256).",
            "results block: prefill=23.2 decode=1.33 tokens/s"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 2.0,
            "init_s": 12.98185,
            "prefill_tokens": 231.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 23.2,
          "provenance": "measured",
          "runs": true,
          "total_ops": 1493,
          "ttft_ms": 10710.0
        },
        "source": "data/device_runs/0.16.0/2026-09-05/shieldstral-1.0-3b-vision-int4-gpu__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-08-28",
          "decode_tokens_per_s": 11.73,
          "delegated_ops": 1187,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-cl-pinned",
            "os_build": "Android 16",
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "VERBOSE: Replacing 1187 out of 1187 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_2048).",
            "VERBOSE: Replacing 1187 out of 1187 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_1024).",
            "VERBOSE: Replacing 1187 out of 1187 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_512).",
            "VERBOSE: Replacing 1187 out of 1187 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_256).",
            "results block: prefill=314.18 decode=11.73 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": 2.0,
            "init_s": 10.11556,
            "prefill_tokens": 231.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 314.18,
          "provenance": "measured",
          "runs": true,
          "total_ops": 1187,
          "ttft_ms": 820.0
        },
        "source": "data/device_runs/0.16.0/2026-08-28/shieldstral-1.0-3b-vision-int4-gpu__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "shieldstral",
    "id": "shieldstral-1.0-3b-vision-int4-gpu",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/mistralai/Shieldstral-1.0-3B",
    "task": "image-text-to-text"
  },
  "pitfalls": [
    "The pre-rebundle published file refused GPU engine creation outright: STABLEHLO_COMPOSITE odml.softmax at 52/1187 ops on the first subgraph (S4 idx 034). The rebundled decoder is composite-free and delegates 1187/1187 x13 subgraphs on the Galaxy S26 — the text sibling's numbers exactly (s5_gate results, device_runs 2026-08-28).",
    "The card's claim 'the vision bundle can replace the text one' was false on GPU for the old file and true again only after this rebundle — the sentence is true only of the _int4_gpu file; the old _int4 stays CPU-only (CARD_ADDITIONS.md §1).",
    "Safety classifier usage: the model scores via a single-token output contract — use the scoring API pattern, not free generation, and letterboxing beats naive resize for image inputs (memory shieldstral-3b-shipped; published HF card)."
  ],
  "schema_version": "1.2"
}