{
  "artifacts": [
    {
      "file": "ema_step_multi_fp32.tflite",
      "sha256": "71b98f74248127f25f6bd0f236f5c99cac729595a044be7a14701be76c091a5b",
      "size_mb": 14.608
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "TODO — the Hub card states the tool and the runtime but publishes no conversion command or script (the repo ships the graphs, config.json and the host loop ema_host.py)",
    "quantization": "none: fp32 weights (card 'Files' / 'Weights': DiT steps fp32 — fp16 weights on the DiT failed the parity bar on one of six sentences)",
    "tool": "litert-torch",
    "tool_version": "0.9.4 (LiteRT 2.2.0 CompiledModel for the checks)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "SQUARED_DIFFERENCE",
      "LOGISTIC"
    ],
    "coverage_ops_pct": 96.4,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 2555,
      "unmatched": 95
    },
    "partitions": 100
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the shipped ema_step_multi_fp32.tflite (15317440 B, sha256 71b98f74248127f25f6bd0f236f5c99cac729595a044be7a14701be76c091a5b = config.json = the Hub LFS sha256; the phone copy digested MATCH /), signatures ema_step_T48, ema_step_T96, ema_step_T192, ema_step_T384, ema_step_T512",
            "latency_p50_ms = the check run's warm median (no rested timing run exists for this cell): (mode generic, run p3 (readme2 / readme3 rows)) summary.warm_median_write_run_read_ms 14.2302605 ms over 2 warm rows of 2 (min 12.916823, max 15.543698, first call 15.543698); compile_ms 34.164896; graph ema_step_multi_fp32.tflite (15317440 B)",
            "10-11 20:32:58.317 27622 27642 I tflite  : Replacing 530 out of 530 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 3 (ema_step_T384).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 347.04, min_corr 1.000000, max_abs 0, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme2__output_0: corr 1.000000, SNR 347.2 dB, max |d| 0; readme3__output_0: corr 1.000000, SNR 347.0 dB, max |d| 0",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph (ai-edge-litert 2.2.0 XNNPACK) on the same inputs: max |d| = 0 <= 1e-5 on every (row, output), so the absolute term holds at every element; max_rel_diff 0.0 (bit-exact)"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 14.2302605,
          "loads": true,
          "max_abs_diff": 0.0,
          "max_rel_diff": 0.0,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 34.164896,
            "gate_leg_first_call_ms": 15.543698,
            "gate_leg_rows": 2.0,
            "gate_leg_warm_max_ms": 15.543698,
            "gate_leg_warm_median_ms": 14.2302605,
            "gate_leg_warm_min_ms": 12.916823,
            "gate_leg_warm_rows": 2.0,
            "max_abs_vs_mac_fp32": 0.0,
            "min_corr_vs_mac_fp32": 0.9999999999999999,
            "min_snr_db_vs_mac_fp32": 347.036,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "threads": 4.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T384 (the shipped file); CPU XNNPACK 4 threads; gate app",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the single-bucket fp32 export ema_step_T96_fp32.tflite (14705024 B, sha256 of the phone copy 3730f3403058988d6c052391afbd6b5923d5923d6df0f39c8db47bc4d7099f22 = the Mac file): the same DiT step graph and weights as the shipped multi-signature file's ema_step_T96 signature, one signature per file (export_graphs.py without --multisig); the work verified identical outputs of the two packagings on the Mac",
            "latency_p50_ms = (mode generic_timing, run p2, rested phone): timing.t_step.per_call_ms_write_run_read median 2.472969 ms (min 2.310886, max 3.01724, n 20) after 3 warm-up calls on fixture row 0 of ema_step_T96.json; run-only median 2.4420835 ms; compile_ms 17.764583; graph ema_step_T96_fp32.tflite (14705024 B); litert 2.2.0, device SM-S942Q, android API 36, threads 4",
            "10-11 19:56:30.499 25318 25338 I tflite  : Replacing 530 out of 530 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 341.10, min_corr 1.000000, max_abs 0, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1__output_0: corr 1.000000, SNR 341.9 dB, max |d| 0; own1__output_0: corr 1.000000, SNR 341.1 dB, max |d| 0",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph (ai-edge-litert 2.2.0 XNNPACK) on the same inputs: max |d| = 0 <= 1e-5 on every (row, output), so the absolute term holds at every element; max_rel_diff 0.0 (bit-exact)"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 2.472969,
          "loads": true,
          "max_abs_diff": 0.0,
          "max_rel_diff": 0.0,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 17.764583,
            "max_abs_vs_mac_fp32": 0.0,
            "min_corr_vs_mac_fp32": 1.0,
            "min_snr_db_vs_mac_fp32": 341.1,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 3.01724,
            "rested_leg_median_ms": 2.472969,
            "rested_leg_min_ms": 2.310886,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 3.0,
            "run_only_median_ms": 2.4420835,
            "threads": 4.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (single-bucket fp32 export ema_step_T96_fp32.tflite); CPU XNNPACK 4 threads; gate app",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the shipped ema_step_multi_fp32.tflite (15317440 B, sha256 71b98f74248127f25f6bd0f236f5c99cac729595a044be7a14701be76c091a5b = config.json = the Hub LFS sha256; the phone copy digested MATCH /), signatures ema_step_T48, ema_step_T96, ema_step_T192, ema_step_T384, ema_step_T512",
            "latency_p50_ms = (mode generic_timing, run p3, rested phone): timing.t_step.per_call_ms_write_run_read median 2.5629429999999997 ms (min 2.319479, max 4.430417, n 20) after 3 warm-up calls on fixture row 0 of ema_step_T96.json; run-only median 2.5363805 ms; compile_ms 35.056718; graph ema_step_multi_fp32.tflite (15317440 B); litert 2.2.0, device SM-S942Q, android API 36, threads 4",
            "check run (mode generic, run p3): 2 fixture rows, warm median 3.8780469999999996 ms (a parity run, not the timing figure); compile_ms 31.877396",
            "10-11 20:35:13.793 31943 31965 I tflite  : Replacing 530 out of 530 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 1 (ema_step_T96).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 341.10, min_corr 1.000000, max_abs 0, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1__output_0: corr 1.000000, SNR 341.9 dB, max |d| 0; own1__output_0: corr 1.000000, SNR 341.1 dB, max |d| 0",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph (ai-edge-litert 2.2.0 XNNPACK) on the same inputs: max |d| = 0 <= 1e-5 on every (row, output), so the absolute term holds at every element; max_rel_diff 0.0 (bit-exact)"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 2.5629429999999997,
          "loads": true,
          "max_abs_diff": 0.0,
          "max_rel_diff": 0.0,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 31.877396,
            "gate_leg_first_call_ms": 5.250677,
            "gate_leg_rows": 2.0,
            "gate_leg_warm_max_ms": 5.250677,
            "gate_leg_warm_median_ms": 3.8780469999999996,
            "gate_leg_warm_min_ms": 2.505417,
            "gate_leg_warm_rows": 2.0,
            "max_abs_vs_mac_fp32": 0.0,
            "min_corr_vs_mac_fp32": 1.0,
            "min_snr_db_vs_mac_fp32": 341.1,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 4.430417,
            "rested_leg_median_ms": 2.5629429999999997,
            "rested_leg_min_ms": 2.319479,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 3.0,
            "run_only_median_ms": 2.5363805,
            "threads": 4.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (the shipped file); CPU XNNPACK 4 threads; gate app",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the single-bucket fp32 export ema_step_T96_fp32.tflite (14705024 B, sha256 of the phone copy 3730f3403058988d6c052391afbd6b5923d5923d6df0f39c8db47bc4d7099f22 = the Mac file): the same DiT step graph and weights as the shipped multi-signature file's ema_step_T96 signature, one signature per file (export_graphs.py without --multisig); the work verified identical outputs of the two packagings on the Mac",
            "latency_p50_ms = (mode generic_timing, run p2, rested phone): timing.t_step.per_call_ms_write_run_read median 2.4051825 ms (min 2.335573, max 3.474011, n 20) after 3 warm-up calls on fixture row 0 of ema_step_T96.json; run-only median 0.09598999999999999 ms; compile_ms 374.07875; graph ema_step_T96_fp32.tflite (14705024 B); litert 2.2.0, device SM-S942Q, android API 36",
            "10-11 20:00:48.878 29869 29890 I tflite  : Replacing 530 out of 530 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 31.12, min_corr 0.999614, max_abs 0.668, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1__output_0: corr 0.999857, SNR 35.4 dB, max |d| 0.41; own1__output_0: corr 0.999614, SNR 31.1 dB, max |d| 0.668",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph on the same inputs: provably violated — max |d| = 0.668 with the row's SNR below 60 dB bounds sum ref^2 so tightly that no element can carry that difference under the rule (the bound: d_i^2 <= 1e-10 + 1e-6 ref_i^2 summed, with sum d^2 = sum ref^2 * 10^(-snr/10)); max_rel_diff not computed. The pass bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass",
            "one step at default precision sits on the pass bar (31.1 dB); four unrolled steps in fp16 read 21.0 dB — the published assignment keeps the steps at fp32 or on the HTP"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 2.4051825,
          "loads": true,
          "max_abs_diff": 0.6677532196044922,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 374.07875,
            "max_abs_vs_mac_fp32": 0.6677532196044922,
            "min_corr_vs_mac_fp32": 0.9996136885633422,
            "min_snr_db_vs_mac_fp32": 31.12,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 3.474011,
            "rested_leg_median_ms": 2.4051825,
            "rested_leg_min_ms": 2.335573,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 3.0,
            "run_only_median_ms": 0.09598999999999999
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (single-bucket fp32 export ema_step_T96_fp32.tflite); GPU default precision (fp16 activations); gate app",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-npu-runner",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the C-API runner ema_npu_runner (LiteRT 2.2.0 C API, --accel npu+cpu --burst 1, the Qualcomm compiler plugin / dispatch and QNN HTP libraries staged in /data/local/tmp/ema_npu, JIT with a new cache dir per run; measured interval = write inputs + run + read outputs per round)",
            "file measured: the single-bucket fp32 export ema_step_T96_fp32.tflite (14705024 B, sha256 of the phone copy 3730f3403058988d6c052391afbd6b5923d5923d6df0f39c8db47bc4d7099f22 = the Mac file): the same DiT step graph and weights as the shipped multi-signature file's ema_step_T96 signature, one signature per file (export_graphs.py without --multisig); the work verified identical outputs of the two packagings on the Mac",
            "latency_p50_ms =: call_ms median 3.6925 ms (min 3.517, max 3.987) over 20 rounds after 5 warm-ups; compile 725.6 ms; the gate app's rested median on the same file is 3.21 ms",
            "VERBOSE: Replacing 530 out of 530 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 110.23, min_corr 1.000000, max_abs 4.7e-05, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1: corr 1.000000, SNR 113.6 dB, max |d| 2.3e-05; own1: corr 1.000000, SNR 110.2 dB, max |d| 4.7e-05",
            "output_match null: the repo element rule cannot be settled from the scorer's per-row summary (max |d| = 4.7e-05 > 1e-5 and the SNR leaves room for the relative term); the bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 3.6925,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 725.6,
            "max_abs_vs_mac_fp32": 4.696846008300781e-05,
            "min_corr_vs_mac_fp32": 0.9999999999952637,
            "min_snr_db_vs_mac_fp32": 110.235,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 3.987,
            "rested_leg_median_ms": 3.6925,
            "rested_leg_min_ms": 3.517,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 5.0,
            "run_only_median_ms": 0.445
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (single-bucket fp32 export ema_step_T96_fp32.tflite); GpuOptions precision=FP32 through the C-API runner (the A/B control of the HTP runs); C-API runner",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the single-bucket fp32 export ema_step_T96_fp32.tflite (14705024 B, sha256 of the phone copy 3730f3403058988d6c052391afbd6b5923d5923d6df0f39c8db47bc4d7099f22 = the Mac file): the same DiT step graph and weights as the shipped multi-signature file's ema_step_T96 signature, one signature per file (export_graphs.py without --multisig); the work verified identical outputs of the two packagings on the Mac",
            "latency_p50_ms = (mode generic_timing, run p2, rested phone): timing.t_step.per_call_ms_write_run_read median 3.2139325000000003 ms (min 2.931406, max 3.749167, n 20) after 3 warm-up calls on fixture row 0 of ema_step_T96.json; run-only median 0.0990365 ms; compile_ms 340.930782; graph ema_step_T96_fp32.tflite (14705024 B); litert 2.2.0, device SM-S942Q, android API 36",
            "10-11 20:00:21.315 29333 29397 I tflite  : Replacing 530 out of 530 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 111.91, min_corr 1.000000, max_abs 3.45e-05, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1__output_0: corr 1.000000, SNR 114.1 dB, max |d| 3.05e-05; own1__output_0: corr 1.000000, SNR 111.9 dB, max |d| 3.45e-05",
            "output_match null: the repo element rule cannot be settled from the scorer's per-row summary (max |d| = 3.45e-05 > 1e-5 and the SNR leaves room for the relative term); the bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 3.2139325000000003,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 340.930782,
            "max_abs_vs_mac_fp32": 3.445148468017578e-05,
            "min_corr_vs_mac_fp32": 0.9999999999967756,
            "min_snr_db_vs_mac_fp32": 111.906,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 3.749167,
            "rested_leg_median_ms": 3.2139325000000003,
            "rested_leg_min_ms": 2.931406,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 3.0,
            "run_only_median_ms": 0.0990365
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (single-bucket fp32 export ema_step_T96_fp32.tflite); GpuOptions precision=FP32; gate app",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 530,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-npu-runner",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": "Qualcomm QNN HTP runtime libraries staged beside the runner (libQnnHtp.so, libQnnHtpPrepare.so, libQnnHtpV81Skel.so / Stub.so, libQnnSystem.so; sha256 MATCH) with LiteRT's libLiteRtCompilerPlugin_Qualcomm.so / libLiteRtDispatch_Qualcomm.so; no version string is printed by the runner"
          },
          "error": null,
          "evidence": [
            "harness: the C-API runner ema_npu_runner (LiteRT 2.2.0 C API, --accel npu+cpu --burst 1, the Qualcomm compiler plugin / dispatch and QNN HTP libraries staged in /data/local/tmp/ema_npu, JIT with a new cache dir per run; measured interval = write inputs + run + read outputs per round)",
            "file measured: the single-bucket fp32 export ema_step_T96_fp32.tflite (14705024 B, sha256 of the phone copy 3730f3403058988d6c052391afbd6b5923d5923d6df0f39c8db47bc4d7099f22 = the Mac file): the same DiT step graph and weights as the shipped multi-signature file's ema_step_T96 signature, one signature per file (export_graphs.py without --multisig); the work verified identical outputs of the two packagings on the Mac",
            "latency_p50_ms =: call_ms (write inputs + run + read outputs) median 1.557 ms (min 1.528, max 1.622) over 20 rounds after 5 warm-ups (runner summary line: {\"n\": \"20\", \"median\": \"1.557\", \"min\": \"1.528\", \"max\": \"1.622\", \"nonfinite_rounds\": \"0\", \"compile_ms\": \"1052.9\", \"fully_accelerated\": \"true\"}); run-only median 1.553 ms; JIT compile with a cold cache 1052.9 ms",
            "INFO: [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 530 ops, from a total of 530 ops. resulted in 1 partitions.",
            "VERBOSE: Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 41.00, min_corr 0.999960, max_abs 0.235, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1: corr 0.999980, SNR 43.9 dB, max |d| 0.235; own1: corr 0.999960, SNR 41.0 dB, max |d| 0.169",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph on the same inputs: provably violated — max |d| = 0.235 with the row's SNR below 60 dB bounds sum ref^2 so tightly that no element can carry that difference under the rule (the bound: d_i^2 <= 1e-10 + 1e-6 ref_i^2 summed, with sum d^2 = sum ref^2 * 10^(-snr/10)); max_rel_diff not computed. The pass bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass",
            "the C-API runner dumps output 0 only, so the HTP parity rows score output 0 (npu_score.py: 'parity of the dumps against the rows file's expected output 0'); for the text graph that is h [1,L,224] — dur [1,L] is not scored on the HTP",
            "the published assignment's step accelerator: 4 x 1.56 ms per piece"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 1.557,
          "loads": true,
          "max_abs_diff": 0.23508071899414062,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 1052.9,
            "max_abs_vs_mac_fp32": 0.23508071899414062,
            "min_corr_vs_mac_fp32": 0.9999603766772515,
            "min_snr_db_vs_mac_fp32": 40.995,
            "parity_call_median_ms": 1.583,
            "parity_rows": 2.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 1.622,
            "rested_leg_median_ms": 1.557,
            "rested_leg_min_ms": 1.528,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 5.0,
            "run_only_median_ms": 1.553
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (single-bucket fp32 export ema_step_T96_fp32.tflite); HTP JIT via the Qualcomm compiler plugin (BURST, cold cache); C-API runner",
          "total_ops": 530,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__galaxy-s26.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "file measured: the single-bucket fp32 export ema_step_T192_fp32.tflite (14726536 B, sha256 b23439147e5cc7b3a6c17c0b2b5a066d70b7dfec23b7142183ba2f5aba2a0e11 read from the file today; the same graph and weights as the shipped multi-signature file's ema_step_T192 signature, one signature per file; identical outputs verified); harness: bench_mac.py --threads 4 --n 20 (ai-edge-litert 2.2.0 Interpreter, XNNPACK)",
            "'ema_step_T192_fp32.tflite            median     3.97 ms  min     3.86 ms  (n=20, threads=4)' (bench_mac.py --threads 4 --n 20: invoke + output read per call, median and min of 20 after warm-up)",
            "parity of the single-bucket fp32 set vs the package oracle; no same-runtime element comparison in the bench, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 3.97,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "rested_leg_median_ms": 3.97,
            "rested_leg_min_ms": 3.86,
            "rested_leg_n": 20.0,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T192 (single-bucket fp32 export ema_step_T192_fp32.tflite); bench_mac.py median of 20",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "file measured: the single-bucket fp32 export ema_step_T384_fp32.tflite (14769544 B, sha256 305c62900df5c65effb707ecfe1608ee13336ed6f30cad290af76e60b7f5eaba read from the file today; the same graph and weights as the shipped multi-signature file's ema_step_T384 signature, one signature per file; identical outputs verified); harness: bench_mac.py --threads 4 --n 20 (ai-edge-litert 2.2.0 Interpreter, XNNPACK)",
            "'ema_step_T384_fp32.tflite            median     6.98 ms  min     6.85 ms  (n=20, threads=4)' (bench_mac.py --threads 4 --n 20: invoke + output read per call, median and min of 20 after warm-up)",
            "parity of the single-bucket fp32 set vs the package oracle; no same-runtime element comparison in the bench, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 6.98,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "rested_leg_median_ms": 6.98,
            "rested_leg_min_ms": 6.85,
            "rested_leg_n": 20.0,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T384 (single-bucket fp32 export ema_step_T384_fp32.tflite); bench_mac.py median of 20",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "file measured: the single-bucket fp32 export ema_step_T48_fp32.tflite (14694272 B, sha256 02422adf0ccd7788605e0ff7f9916309631eee90063ba945bcdeec81088237b6 read from the file today; the same graph and weights as the shipped multi-signature file's ema_step_T48 signature, one signature per file; identical outputs verified); harness: bench_mac.py --threads 4 --n 20 (ai-edge-litert 2.2.0 Interpreter, XNNPACK)",
            "'ema_step_T48_fp32.tflite             median     1.97 ms  min     1.84 ms  (n=20, threads=4)' (bench_mac.py --threads 4 --n 20: invoke + output read per call, median and min of 20 after warm-up)",
            "parity of the single-bucket fp32 set vs the package oracle; no same-runtime element comparison in the bench, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 1.97,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "rested_leg_median_ms": 1.97,
            "rested_leg_min_ms": 1.84,
            "rested_leg_n": 20.0,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T48 (single-bucket fp32 export ema_step_T48_fp32.tflite); bench_mac.py median of 20",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "file measured: the single-bucket fp32 export ema_step_T512_fp32.tflite (14798216 B, sha256 4b02bf0545a3db19688ed9d2ee9c2eec87bffd1e57ad61e640dcbd9b1bb7c8cc read from the file today; the same graph and weights as the shipped multi-signature file's ema_step_T512 signature, one signature per file; identical outputs verified); harness: bench_mac.py --threads 4 --n 20 (ai-edge-litert 2.2.0 Interpreter, XNNPACK)",
            "'ema_step_T512_fp32.tflite            median     9.57 ms  min     9.38 ms  (n=20, threads=4)' (bench_mac.py --threads 4 --n 20: invoke + output read per call, median and min of 20 after warm-up)",
            "parity of the single-bucket fp32 set vs the package oracle; no same-runtime element comparison in the bench, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 9.57,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "rested_leg_median_ms": 9.57,
            "rested_leg_min_ms": 9.38,
            "rested_leg_n": 20.0,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T512 (single-bucket fp32 export ema_step_T512_fp32.tflite); bench_mac.py median of 20",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "file measured: the single-bucket fp32 export ema_step_T96_fp32.tflite (14705024 B, sha256 3730f3403058988d6c052391afbd6b5923d5923d6df0f39c8db47bc4d7099f22 read from the file today; the same graph and weights as the shipped multi-signature file's ema_step_T96 signature, one signature per file; identical outputs verified); harness: bench_mac.py --threads 4 --n 20 (ai-edge-litert 2.2.0 Interpreter, XNNPACK)",
            "'ema_step_T96_fp32.tflite             median     2.76 ms  min     2.49 ms  (n=20, threads=4)' (bench_mac.py --threads 4 --n 20: invoke + output read per call, median and min of 20 after warm-up)",
            "parity of the single-bucket fp32 set vs the package oracle; no same-runtime element comparison in the bench, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 2.76,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "rested_leg_median_ms": 2.76,
            "rested_leg_min_ms": 2.49,
            "rested_leg_n": 20.0,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T96 (single-bucket fp32 export ema_step_T96_fp32.tflite); bench_mac.py median of 20",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the host loop hostloop_e2e.py --split --multi --threads 4 (ai-edge-litert 2.2.0 Interpreter, XNNPACK 4 threads; normalizer-tr frontend + the package's chunker; per-stage wall time of single calls: write inputs + run + read outputs) on the shipped multi-signature files (multi files, sha256 = config.json = the Hub LFS)",
            "Per text the piece's L / T, the bucket Lb / Tb it ran in, text_ms, sound_ms (= align + the four steps + the host Euler mixing), decode_ms per window, first_audio_ms, total_ms; forms {\"text\": \"fp32\", \"sound\": \"fp32\", \"decoder\": \"fp16\"}, split True, threads 4",
            "(verify_tflite.py --multi on the same files vs the package oracle fixtures/): per fixture frames_ok, |dh|, |ddur|, |dlat|, corr_lat, |daudio|, corr_audio, snr_db — readme1: frames_ok True, |dh| 1.91e-06, |ddur| 1.31e-06, corr_lat 1.000000000, SNR 56.40 dB, corr_audio 0.9999995; readme2: frames_ok True, |dh| 1.79e-06, |ddur| 2.62e-06, corr_lat 1.000000000, SNR 56.82 dB, corr_audio 0.9999995; readme3: frames_ok True, |dh| 1.43e-06, |ddur| 2.38e-06, corr_lat 1.000000000, SNR 56.84 dB, corr_audio 0.9999996; own1: frames_ok True, |dh| 1.19e-06, |ddur| 1.31e-06, corr_lat 1.000000000, SNR 57.11 dB, corr_audio 0.9999996; own2: frames_ok True, |dh| 1.43e-06, |ddur| 1.55e-06, corr_lat 1.000000000, SNR 56.57 dB, corr_audio 0.9999995; own3: frames_ok True, |dh| 1.31e-06, |ddur| 1.79e-06, corr_lat 1.000000000, SNR 56.57 dB, corr_audio 0.9999995",
            "output_match null: the comparison is against the upstream PyTorch package (the oracle), not a same-runtime CPU reference (shape); the pass bar passes on all six fixtures"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "align_plus_4_steps_ms_own1_L64_T96": 10.7476,
            "align_plus_4_steps_ms_own2_L128_T192": 15.7549,
            "align_plus_4_steps_ms_own3_L128_T192": 15.5673,
            "align_plus_4_steps_ms_readme1_L64_T96": 10.8862,
            "align_plus_4_steps_ms_readme2_L256_T384": 27.7937,
            "align_plus_4_steps_ms_readme3_L256_T384": 27.7448,
            "corr_latents_vs_pytorch_own1": 0.9999999999519535,
            "corr_latents_vs_pytorch_own2": 0.9999999999510488,
            "corr_latents_vs_pytorch_own3": 0.9999999998416187,
            "corr_latents_vs_pytorch_readme1": 0.999999999984055,
            "corr_latents_vs_pytorch_readme2": 0.9999999996144746,
            "corr_latents_vs_pytorch_readme3": 0.9999999997009323,
            "max_abs_dlat_vs_pytorch_own1": 0.00022876262664794922,
            "max_abs_dlat_vs_pytorch_own2": 0.0002440810203552246,
            "max_abs_dlat_vs_pytorch_own3": 0.0007377862930297852,
            "max_abs_dlat_vs_pytorch_readme1": 0.00012774765491485596,
            "max_abs_dlat_vs_pytorch_readme2": 0.00072479248046875,
            "max_abs_dlat_vs_pytorch_readme3": 0.0008115768432617188,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_step_T{48,96,192,384,512} (the shipped file); host loop, one call per fixture",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_step_multi_fp32__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "ema-lightning",
    "id": "ema-lightning__ema_step_multi_fp32",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/EMA-Lightning-LiteRT",
    "task": "text-to-speech"
  },
  "pitfalls": [
    "Step contract (card 'Files'): x [1,T,64], cond [1,T,224], bc [1,6,224], sg [1,2,224], fbias [1,1,1,T] -> v [1,T,64] — one DiT velocity pass, four calls per piece; signatures ema_step_T{48,96,192,384,512}, shapes fixed per signature. bc / sg come from ema_step_table.npz (the four steps' modulation vectors at t = 0, 1/4, 1/2, 3/4) and the sampler is the host's: x1 = x + (1 - t)·v, then x = (1 - t')·noise' + t'·x1 with seeded Gaussian noise (the package's DMD2-distilled sampler; seed 0 on CPU reproduces the PyTorch package) (card 'How the host loop works' step 4).",
    "Run the steps at fp32: fp16 weights on the DiT failed the bar on one of six sentences (26–28 dB) and GPU fp16 activations sit on the bar for one step (31 dB) and fail for the four unrolled steps (21 dB); dynamic int8 weights fail everywhere (card 'Files' / 'Weights', 'Parity').",
    "The HTP (LiteRT's NPU accelerator with the Qualcomm compiler plugin and runtime libraries, JIT) is this graph's fastest home on the Galaxy S26: 1.56 ms per step at 41 dB, compile 1 s per process — the card's fastest configuration runs the four steps there; without the vendor NPU libraries the steps run on the CPU (2.5 ms each) (card 'Performance').",
    "Turkish only, one voice; fixed buckets — a piece longer than 256 letters / 512 frames is cut by the chunker first, speeds far from 1.0 may move a piece to a larger bucket (card 'Limitations')."
  ],
  "schema_version": "1.2"
}
