{
  "_comment": [
    "MedASR per-tensor tolerances for compare_tensors.py.",
    "",
    "CORRECTNESS REGIME",
    "- Reference: google/medasr @ ae1e4845b4b07479735d93e1e591e566435b7104",
    "  loaded via transformers v5.0.0.dev0 (commit 65dc261512cbdb1ee72b88ae5b222f2605aad8e5,",
    "  AutoModelForCTC + AutoProcessor, attn_implementation=eager, dtype=float32).",
    "  Dumper: scripts/dump_reference_medasr_transformers.py.",
    "- C++: ggml CPU fp32 compute, weights stored as F32 in GGUF, --backend cpu --threads 1.",
    "- KV cache dtype: n/a (CTC head, no autoregressive decoder).",
    "- Mel frontend: PRODUCTION C++ MedAsrMelFrontend (LasrFeatureExtractor",
    "  port: manual unfold + rfft, Hann symmetric, no preemph, no dither,",
    "  HTK 128-band kaldi mel, log(clamp(power, 1e-5)), center=False).",
    "  No env-var injection; mel.in measures the C++ vs reference frontend",
    "  delta directly, folding Step 7 frontend parity into the same regime.",
    "",
    "FINALIZED FROM OBSERVED C++ DRIFT (Stage 4)",
    "Recipe: finalized = max(1.5 * observed, 1e-6). Every entry tightened",
    "vs the Stage 2 provisional magnitude-budget floor (the C++ drift is",
    "consistently 30 to 300 times below the 1e-4 * p99_abs ceiling).",
    "",
    "DOMINANT DRIFT SOURCES",
    "- mel.in (max ~5e-4): C++ rfft accumulates in fp32 throughout; the",
    "  reference LasrFeatureExtractor's mel-filterbank multiplication is",
    "  computed in fp64 before the log. log(clamp) compresses this back to",
    "  small absolute drift on log-mel values that sit in roughly [-12, 6].",
    "- mul_mat accumulation order. ggml's CPU mul_mat path accumulates in",
    "  a different sum-order than torch's eager Linear (the latter dispatches",
    "  to MKL/Accelerate). For 512-deep dot products of normally-distributed",
    "  activations the order difference is ~1e-6 relative per element, which",
    "  compounds modestly through the 17-layer residual stream.",
    "- The macaron residual scalars [1.5, 0.5] and the conv residual scalars",
    "  [2.0, 1.0] amplify the running residual within each block (up to ~2e6",
    "  raw magnitude on enc.block.0.post_ff2 before the per-block norm_out",
    "  collapses it). post_* drift budgets scale to that magnitude; block.<i>.out",
    "  budgets are post-LN so back near 1.",
    "- enc.subsampling.out absorbs both the mel drift and the post-ReLU",
    "  conv-1d accumulation order at d_model=512.",
    "",
    "NOT REGIMENED HERE",
    "- Quantized GGUF tolerances (Stage 6/7 reports).",
    "- F16 KV cache (n/a here).",
    "- Variable-length batch parity beyond bs=1 (Stage 4 Step 9 gate)."
  ],
  "mel.in": {
    "max_abs": 0.000784,
    "mean_abs": 2.47e-06,
    "_seen_in": ["medasr"]
  },
  "enc.subsampling.out": {
    "max_abs": 0.328,
    "mean_abs": 0.0250,
    "_seen_in": ["medasr"]
  },
  "enc.block.0.post_ff1": {
    "max_abs": 0.422,
    "mean_abs": 0.0382,
    "_seen_in": ["medasr"]
  },
  "enc.block.0.post_attn": {
    "max_abs": 0.388,
    "mean_abs": 0.0386,
    "_seen_in": ["medasr"]
  },
  "enc.block.0.post_conv": {
    "max_abs": 0.803,
    "mean_abs": 0.0788,
    "_seen_in": ["medasr"]
  },
  "enc.block.0.post_ff2": {
    "max_abs": 1.546,
    "mean_abs": 0.1274,
    "_seen_in": ["medasr"]
  },
  "enc.block.0.out": {
    "max_abs": 0.000641,
    "mean_abs": 1.06e-05,
    "_seen_in": ["medasr"]
  },
  "enc.block.7.out": {
    "max_abs": 0.000824,
    "mean_abs": 1.16e-05,
    "_seen_in": ["medasr"]
  },
  "enc.block.8.out": {
    "max_abs": 0.000732,
    "mean_abs": 1.17e-05,
    "_seen_in": ["medasr"]
  },
  "enc.block.16.out": {
    "max_abs": 0.000280,
    "mean_abs": 4.53e-06,
    "_seen_in": ["medasr"]
  },
  "enc.out_norm.out": {
    "max_abs": 4.65e-06,
    "mean_abs": 1e-06,
    "_seen_in": ["medasr"]
  },
  "enc.ctc_logits": {
    "max_abs": 5.72e-05,
    "mean_abs": 3.41e-06,
    "_seen_in": ["medasr"]
  }
}