- Simplified estimator count retrieval in CosyVoice and CosyVoice2 classes to directly access the configs dictionary.
- Adjusted memory pool limit in the ONNX to TensorRT conversion function from 8GB to 1GB for optimized resource management.
- Removed the queue-based stream pool and integrated direct CUDA stream usage for improved performance.
- Simplified inference methods by eliminating unnecessary synchronization and stream management code.
- Enhanced logging for better tracking of synthesis operations and performance metrics.
- Updated the model class to support CUDA stream context management, ensuring efficient resource utilization during inference.
- Introduced a queue-based system for managing CUDA streams to improve inference performance.
- Updated inference methods to utilize CUDA streams for asynchronous processing.
- Added an EstimatorWrapper class to manage TensorRT estimators, allowing for efficient execution context handling.
- Modified model loading functions to support estimator count configuration.
- Improved logging and performance tracking during inference operations.