fix: use eager attention for SDPA compatibility with transformers >=4.36

[majiayu000]

Summary

• Set attn_implementation='eager' when creating LlamaConfig
• Ensures output_attentions=True works correctly

Problem

Fixes #339

In transformers >=4.36, SDPA became the default attention implementation. However, SDPA doesn't support output_attentions=True, which Chatterbox uses during inference with voice references.

This causes:

ValueError: The `output_attentions` attribute is not supported when using the `attn_implementation` set to sdpa.

Solution

Explicitly set attn_implementation='eager' in LlamaConfig. Eager attention fully supports all features including output_attentions.

Impact

• Voice cloning and voice conversion features now work with modern transformers versions
• No performance regression for typical use cases (inference already uses output_attentions)

[George0828Zhang]

Looking at this comment

chatterbox/src/chatterbox/models/t3/inference/alignment_stream_analyzer.py

Line 60 in ed27b95

# using it for all layers slows things down too much. We can apply it to just one layer

It is likely intended for a single layer. Is it possible to only apply eager to that layer only, and keep fast sdpa kernel on the other layers?

[majiayu000]

@George0828Zhang Thanks for the great suggestion!

I've updated the implementation to apply eager attention only to the specific layers that need output_attentions=True.

Key changes:

• Removed global attn_implementation='eager' from LlamaConfig
• Modified _add_attention_spy() to wrap the forward method of layers 9, 12, 13 (based on LLAMA_ALIGNED_HEADS)
• Each wrapped layer temporarily switches to eager attention, calls forward with output_attentions=True, then restores the original implementation

This preserves SDPA performance for all other layers while supporting alignment stream analysis.

Please take a look and let me know if this addresses your concern!

[George0828Zhang]

Thanks. I installed your commit and it seemed to work fine.

[majiayu000]

Happy to help!