{
  "artifacts": [
    {
      "file": "nemotron3_diar_frontend.tflite",
      "sha256": "c01df58c31b425183c5a008552b517e53007fee39748baa2a4c49c1a55072810",
      "size_mb": 2.001
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-09-25",
        "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": 0.425,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-09-25",
        "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": 0.532,
        "loads": true,
        "max_rel_diff": 0.0013167669386284582,
        "output_match": true,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-09-25/nemotron-3-diarization__nemotron3_diar_frontend.json"
  },
  "conversion": {
    "command": "python build_nemotron3diar.py --run-dir $RUN --model-dir $RUN/hf_model --mode low_latency --ln safe  (graph A = exports/n3d_frontend.tflite, renamed on publish)",
    "quantization": "none — float32 (graph A is published as the float32 export; conversion/README.md step 6)",
    "tool": "litert-torch",
    "tool_version": "0.9.4 (litert-converter 0.4.0; torch 2.11.0; ai-edge-quantizer 0.9.0 for the float16 cast)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [],
    "coverage_ops_pct": 100.0,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 2
    },
    "partitions": 1
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-09-24",
          "decode_tokens_per_s": null,
          "delegated_ops": 2,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-n3d-selftest",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "nemotron3diar SelfTest (zoo main a7ce147, HF android/SelfTest.kt): one Environment, CompiledModel(Accelerator.GPU) per graph; inputs = 7 captured steps of the 97.6 s upstream example clip, fed with the transformers fp32 reference's own encoder inputs; latency = write inputs -> run -> readFloat of every output, median of 20 timed runs after 5 warm-ups at step 128 (L=541 rows); parity = graph output vs the reference's chunk logits over the L*8 real rows; sources ~/code/codex-conversions/2026-09-24/n3d-litert/results/gate_device.json, gate_device_offline.json, device_logcat_<tag>.txt (\"Replacing N out of N node(s) with delegate (LITERT_CL) node, yielding 1 partitions\")",
            "run tag safe_fp16_fp32; GPU precision FP32; conditions at start: screen_interactive=True keyguard_locked=False plugged=usb battery=100% 33C thermal_status=0 thermal_headroom=0.673; thermal after 0",
            "graph A output rows vs the reference's packed embedding rows (|x| up to 141): max|d| 2.0e-4 at FP32 (supervisor recompute over the 7 steps from device/safe_fp16_fp32/n3d/outA_step*.bin); the app always runs graph A at FP32 because its rows persist in the speaker cache"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 0.2066925,
          "loads": true,
          "max_abs_diff": 0.000204,
          "max_rel_diff": null,
          "metrics": {
            "iterations": 20,
            "latency_max_ms": 0.706146,
            "latency_min_ms": 0.161771,
            "load_compile_ms": 81.659114
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32",
          "total_ops": 2,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-09-24/nemotron-3-diarization__nemotron3_diar_frontend__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-09-24",
          "decode_tokens_per_s": null,
          "delegated_ops": 2,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-n3d-selftest",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "nemotron3diar SelfTest (zoo main a7ce147, HF android/SelfTest.kt): one Environment, CompiledModel(Accelerator.GPU) per graph; inputs = 7 captured steps of the 97.6 s upstream example clip, fed with the transformers fp32 reference's own encoder inputs; latency = write inputs -> run -> readFloat of every output, median of 20 timed runs after 5 warm-ups at step 128 (L=541 rows); parity = graph output vs the reference's chunk logits over the L*8 real rows; sources ~/code/codex-conversions/2026-09-24/n3d-litert/results/gate_device.json, gate_device_offline.json, device_logcat_<tag>.txt (\"Replacing N out of N node(s) with delegate (LITERT_CL) node, yielding 1 partitions\")",
            "run tag safe_fp16_default; GPU precision default (fp16 compute); conditions at start: screen_interactive=True keyguard_locked=False plugged=usb battery=100% 33C thermal_status=0 thermal_headroom=0.623; thermal after 0",
            "graph A at default precision: max|d| 0.38 vs the reference rows (supervisor recompute from device/safe_fp16_default/n3d/outA_step*.bin) — not used by the app"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 0.2169005,
          "loads": true,
          "max_abs_diff": 0.381,
          "max_rel_diff": null,
          "metrics": {
            "iterations": 20,
            "latency_max_ms": 1.105781,
            "latency_min_ms": 0.171875,
            "load_compile_ms": 112.759218
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=default (fp16 compute)",
          "total_ops": 2,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-09-24/nemotron-3-diarization__nemotron3_diar_frontend__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "nemotron-3-diarization",
    "id": "nemotron-3-diarization__nemotron3_diar_frontend",
    "license": "openmdw-1.1",
    "source_url": "https://huggingface.co/litert-community/Nemotron-3-Diarization-LiteRT",
    "task": "voice-activity-detection"
  },
  "pitfalls": [
    "Run graph A at GPU precision FP32: its rows stay in the speaker cache and FIFO for the whole session; at default precision they differ from the reference by up to 0.38 (|x| <= 141), at FP32 by 2.0e-4, for 0.2 ms (card: Conversion notes, 'Graph A in FP32').",
    "Input is log-mel of one chunk: 104 frames x 128 slaney mel bins, pre-emphasis 0.97, 400-sample Hann in a 512-point FFT, hop 160, log(x + 2^-24), no normalization; zero rows after the last frame (card: Graph interfaces, Streaming loop).",
    "Log-mel must match torch.stft rounding: a float32 radix-2 FFT was 2.7e-4 off in the log domain; the Kotlin host ports pocketfft's real FFT and the Python host uses numpy's float32 path rfft(norm='forward') * 512 (card: Conversion notes, 'The FFT of the host mel')."
  ],
  "schema_version": "1.2"
}