 diodiogod
|
ae2998b67d
|
Apply IndexTTS-2 upstream updates (September 18 → November 6, 2025)
Integrated all 5 upstream commits with custom modifications preserved:
1. Empty generator fix (750d9d9) - Refactored infer() to generator pattern
- Created wrapper infer() that delegates to infer_generator()
- Renamed monolithic method to infer_generator()
- Changed final returns to yield for generator compatibility
- Handles both streaming and non-streaming modes
2. Streaming implementation (b0c6ab8) - Added low-latency streaming
- Added stream_return parameter for generator mode
- Added more_segment_before parameter for streaming segmentation
- Integrated with generator-based inference
3. UNK token warnings (34be9bf) - Token validation and user feedback
- Detects unknown tokens in BPE model
- Warns users with token IDs and suggestions
4. S2Mel optimization (31e7e85) - Mel-spectrogram generation speedup
- Optimized fused_add_tanh_sigmoid_multiply() with torch.split
- Fixed sequence_mask() with max_length parameter
- Added enable_torch_compile() to CFM and MyModel classes
- Added use_torch_compile parameter to infer_v2.py
5. GPT2 acceleration (c1ef414) - KV-cache and FlashAttention optimization
- Created new /accel/ module with 5 files (609-2500 lines)
- Added use_accel parameter to UnifiedVoice
- Integrated AccelInferenceEngine for 1.5-3x speedup
- Optional feature (requires flash-attn library)
Custom modifications preserved:
- QwenEmotion lazy loading with path fallbacks
- W2V-BERT loading from TTS folder (unified downloader integration)
- All custom logging and error handling
- Custom emotion vector processing
Generated upstream tracking documentation in docs/Upstream_Engine_Tracking/
|
2025-11-06 19:32:28 -03:00 |
|