Major fixes:
- Embeddings: use actual checkpoint size (9088 entries with padded codebooks)
- GQA: fix KV head repetition dimension (was interleaving wrong, corrupting
all attention computations in backbone and acoustic transformer)
- Codec decoder: restructure to flat decoder_blocks[0-7] matching checkpoint,
use nn.utils.parametrizations.weight_norm, disable conv bias, fix dtype
- Tokenizer: use official mistral_common SpeechRequest API for correct prompt
format (BOS, BEGIN_AUDIO, AUDIO×N, /INST, text, INST, BEGIN_AUDIO)
- Pipeline: add initial AUDIO token feed before generation loop, fix voice
embedding injection positions
- Loader: handle folder_paths KeyError gracefully
Status: model loads, generates coherent speech on MPS, codec decodes to
24kHz audio. Text conditioning not yet working correctly — model produces
plausible speech but doesn't follow the input text.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
