Disable token timestamps for distilled models detected by config, not name

[gsdali]

Problem

scripts/generate_model.py only disables token-level timestamps for models whose version string contains distil. Some distilled models are not named distil-* yet still ship the original (e.g. large-v3) generation_config, whose alignment_heads reference decoder layers that no longer exist after distillation. Generating token timestamps then crashes with an IndexError in compute_alignment_heads_attention_weights.

Repro: whisperkit-generate-model --model-version kotoba-tech/kotoba-whisper-bilingual-v1.0 — the model has 2 decoder layers but ships large-v3 alignment_heads up to layer 25; the WhisperTextDecoder step fails.

Fix

In addition to the existing name-based check, detect the condition from the config: if any alignment_heads entry references a layer >= decoder_layers, disable token timestamps. Falls back gracefully (warns and continues) if the config can't be read.

Evidence

This change was applied when converting kotoba-tech/kotoba-whisper-bilingual-v1.0 (whisperkittools 0.4.2, coremltools 9.0, torch 2.5.0). With it, the guard fires and the conversion completes:

INFO:scripts.generate_model:Disabling token-level timestamps: alignment_heads reference decoder layers beyond decoder_layers=2
INFO:argmaxtools.test_utils:torch2coreml PSNR=57.4   # TextDecoder
INFO:argmaxtools.test_utils:torch2coreml PSNR=53.9   # AudioEncoder
INFO:argmaxtools.test_utils:torch2coreml PSNR=59.8   # MelSpectrogram

Without the change the same run fails at the decoder step with the IndexError above. I have not re-run the full suite against other models; the change only widens an existing disable condition and is a no-op for models without oversized alignment_heads.

🤖 Generated with Claude Code