{
  "artifacts": [
    {
      "file": "julia1_s1024_fp32.tflite",
      "sha256": "b4f260595776c0a126f0e1aadbefb0c2acc6e0b1ca0cfe9499b4d322d591e632",
      "size_mb": 179.771
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-10-09",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": null,
        "latency_p50_ms": 2014.363,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-10-09",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": true,
        "latency_p50_ms": 43.43,
        "loads": true,
        "max_rel_diff": 0.006589250508776393,
        "output_match": true,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-09/julia-1__julia1_s1024_fp32.json"
  },
  "conversion": {
    "command": "conversion/ scripts of the Hub repo (host-side token lookup, rotary tables baked per window, sliding-window band as a constant, question type as a one-hot matmul, the two head layers written out, exact GELU); REPRODUCE.md of the Hub repo",
    "quantization": "float32 weights: 1,704 operators, no quantization; the token table is the checkpoint's float32 embedding rounded to float16 on the host side (card 'Files', 'Provenance').",
    "tool": "litert-torch",
    "tool_version": "0.9.3 (torch 2.12.1; transformers 5.0.0; ai-edge-litert 2.1.6 for the desktop checks)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "SQUARED_DIFFERENCE"
    ],
    "coverage_ops_pct": 97.1,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1654,
      "unmatched": 50
    },
    "partitions": 51
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-09-30",
          "decode_tokens_per_s": null,
          "delegated_ops": 1704,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-julia-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.julia.gate (android_gate/, Kotlin CompiledModel 2.2.0), one cold process per run, fixture ids + markers pushed as JSON, host token-table lookup outside the timed interval; timed interval = input writes + run + output readback per row.",
            "conditions: before thermal status None, battery 41.3 C; after thermal status None, battery 41.9 C; screen on, USB powered, no cool-start wait.",
            "parity vs the author's implementation (scripts/device_compare.py): same winning option on 100/100 rows (18 boundary = reference top-1 < 0.9); max |dprob| 0.00767; rows over 0.01: 0.",
            "output_match null: no Mac CPU check of these graph bytes holds per-row logits for these ids (keep aggregates only); the argmax / |dprob| metrics against the author's implementation are its own criterion.",
            "09-30 09:51:09.888 30986 31015 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "09-30 09:51:09.994 30986 31015 I tflite  : Replacing 1704 out of 1704 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 298.463958,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "argmax_equal_official": 100,
            "battery_temp_c_end": 41.9,
            "battery_temp_c_start": 41.3,
            "boundary_rows": 18,
            "first_call_ms": 163.107187,
            "fixtures": 100,
            "load_compile_ms": 1473.3325,
            "max_abs_dprob_vs_official": 0.00767,
            "partitions": 1,
            "rows_dprob_gt_1e2": 0,
            "warm_max_ms": 305.159843,
            "warm_median_ms": 298.463958,
            "warm_min_ms": 155.552344
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; host fp16 token table (shipped form); debug gate app com.julia.gate",
          "total_ops": 1704,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-09-30/julia-1__julia1_s1024_fp32__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "julia-1",
    "id": "julia-1__julia1_s1024_fp32",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/Julia-1-LiteRT",
    "task": "text-classification"
  },
  "pitfalls": [
    "Weight storage does not set the GPU compute precision: request GpuOptions(precision = FP32). The fp16 paths change answers: S26 Hexagon NPU (JIT, fp32 graph) 685/706 same argmax with probabilities off by up to 0.418; Mac Metal at its default fp16 precision 375/400 (card 'fp16 and the NPU', 'Measured quality and performance').",
    "The graph takes host-built inputs: inputs_embeds [1,1024,384] from julia1_token_table_fp16.bin ([256000,384] little-endian float16 rows, padding positions included), attention_mask [1,1024] (1 real / 0 pad), qtype_onehot [1,3] (choice / score / noul); output token_logits [1,1024], read at the <mask> marker positions; softmax at temperature 1 runs on the host (card 'Host contract').",
    "Sequence: <bos> '{type} question: {question}' <eos> <mask> ' option0' <mask> ' option1' … <eos> state <eos>; 2 to 20 options, at most 48 tokens each; a request that does not fit raises an error, nothing is truncated (35 of the 2,000 typed-decisions questions need more than 512 tokens and fit the S1024 graph) (card 'Host contract').",
    "The graph computes LayerNorm on x*2^-k with eps*2^-2k (k = 8 in layers 12-21 and the final norm, k = 2 in the head and scorer norms), uses -1e4 attention masks and scales q by 1/8 before QK^T; exact in fp32 (bit-identical token logits), finite but not the same answers in fp16 (card 'fp16 and the NPU', 'Provenance').",
    "The float16 token table changes no answer and moves probabilities by up to 0.0077 (desktop CPU and S26 GPU); the host lookup is outside the measured graph time (Kotlin loop 19.0 ms float32 table / 23.6 ms float16 table per question in the gate app) (card 'Files', 'Measured quality and performance').",
    "Validated against the author's runtime (CPU FP32), not task accuracy; one phone (Galaxy S26 SM-S942Q, LiteRT 2.2.0 Kotlin CompiledModel); English text only; S26 GPU explicit FP32: 100/100 rows (the 35 typed-decisions questions longer than 512 tokens + 65 others) with the shipped float16 table, max probability difference 0.0077, median 298.5 ms at thermal status 2; desktop CPU all 2,000 typed-decisions questions with the author's answers (426/600 choice, 542/800 score, 483/600 noul) (card 'Limits', 'Measured quality and performance')."
  ],
  "schema_version": "1.2"
}
