0.1.2

MOSS-TTS-ComfyUI v0.1.2

Fixed

• Fixed MOSS-Audio-Tokenizer-v2 FlashAttention fallback under ComfyUI/AIMDO dynamic VRAM.
• The codec now prefers the ComfyUI runtime device tag when entering inference autocast, instead of relying on the first parameter device.
• This prevents the codec from incorrectly seeing dtype=torch.float32 and falling back to SDPA when FlashAttention is available.

Improved

• Added a regression test to verify codec autocast uses CUDA BF16 when the model is CPU-staged but registered for CUDA runtime execution.

Validation

• Tested with the ComfyUI virtual environment.
• pytest tests -q
• Result: 11 passed