metadata.json

{
  "model": "canary-1b-v2",
  "window_samples": 240000,
  "precision": "FP16",
  "mel_shape": [
    1,
    128,
    1501
  ],
  "enc_states_shape": [
    1,
    188,
    1024
  ],
  "enc_len": 188,
  "components": {
    "preprocessor": "canary_preprocessor.mlpackage",
    "encoder": "canary_encoder.mlpackage"
  },
  "optimize": {
    "frontend": "preprocessor+encoder (copied from canary_coreml_fp16enc_sliced)",
    "decoder": "kv-cache cross+step (copied from canary_coreml_kv)",
    "deploy": "compute_units=cpu_and_ne",
    "note": "Self-contained KV-cache build: front end (preprocessor + encoder) from canary_coreml_fp16enc_sliced + stateful KV-cache decoder (cross_kv + decoder_kv) from canary_coreml_kv. Run with validate_e2e_kv.py / the Swift KV host."
  },
  "decoder_kv": {
    "cross_file": "canary_cross_kv.mlpackage",
    "step_file": "canary_decoder_kv.mlpackage",
    "l_dec": 238,
    "t_enc": 188,
    "n_layers": 8,
    "n_heads": 8,
    "head_size": 128,
    "max_generation_delta": 50,
    "vocab_size": 16384,
    "pad": 2,
    "eos": 3,
    "seed": [
      16053,
      7,
      4,
      16,
      64,
      64,
      5,
      9,
      11,
      13
    ],
    "kv_cache": true,
    "neg_inf": -10000.0
  }
}