{
  "_comment": [
    "Qwen3-ASR 1.7B per-tensor tolerances for compare_tensors.py.",
    "",
    "Reference: Alibaba qwen_asr v0.0.6 (author repo, transformers host)",
    "running BF16 inference on CPU.",
    "C++: ggml compute (CPU/Metal/Vulkan), weights stored as BF16 in the",
    "accuracy GGUF.",
    "",
    "Dump points follow the audio-llm pattern (final encoder block is",
    "enc.block.23.out for the 1.7B variant, not .17):",
    "  enc.mel.in            - input log-mel after frontend normalization",
    "  enc.subsample.out     - after 3x Conv2d downsampler + flatten + linear",
    "  enc.pos_add.out       - after sinusoidal position add",
    "  enc.block.0.out       - first encoder block output",
    "  enc.block.23.out      - final encoder block output (1.7B has 24 layers)",
    "  enc.ln_post.out       - after final LayerNorm",
    "  enc.proj.out          - after proj1 + GELU + proj2 (d=output_dim)",
    "  dec.token_emb         - LM input embedding pre-injection",
    "  dec.audio_injected    - embedding after audio feature scatter",
    "  dec.block.0.out       - first LM block output",
    "  dec.block.27.out      - final LM block output",
    "  dec.out_before_head   - after final LM RMSNorm, before lm_head",
    "  dec.logits_raw        - pre-softmax next-token logits",
    "  dec.logits            - post-log-softmax (use inf tolerance)",
    "",
    "Budgets are calibrated to ~2x the max observed across CPU, Metal,",
    "and Vulkan on JFK. Vulkan dominates most encoder tensors due to",
    "F16 intermediate precision in its matmul path; bounds honor that",
    "without over-loosening for CPU/Metal regressions."
  ],
  "enc.mel.in": {
    "max_abs": 0.01,
    "mean_abs": 0.001
  },
  "enc.subsample.out": {
    "max_abs": 0.15,
    "mean_abs": 0.01
  },
  "enc.pos_add.out": {
    "max_abs": 0.15,
    "mean_abs": 0.01
  },
  "enc.block.0.out": {
    "max_abs": 4.0,
    "mean_abs": 0.04
  },
  "enc.block.23.out": {
    "max_abs": 65.0,
    "mean_abs": 0.5
  },
  "enc.ln_post.out": {
    "max_abs": 6.0,
    "mean_abs": 0.15
  },
  "enc.proj.out": {
    "max_abs": 0.25,
    "mean_abs": 0.006
  },
  "dec.token_emb": {
    "max_abs": 4e-06,
    "mean_abs": 4e-07
  },
  "dec.audio_injected": {
    "max_abs": 0.25,
    "mean_abs": 0.006
  },
  "dec.block.0.out": {
    "max_abs": 7.0,
    "mean_abs": 0.1
  },
  "dec.block.27.out": {
    "max_abs": 400.0,
    "mean_abs": 8.0
  },
  "dec.out_before_head": {
    "max_abs": 180.0,
    "mean_abs": 0.8
  },
  "dec.logits_raw": {
    "max_abs": 6.0,
    "mean_abs": 1.0
  },
  "dec.logits": {
    "max_abs": "inf",
    "mean_abs": "inf"
  }
}