Release v1.0.2 — RAM Optimization & Storage Cleanup
This update focuses on significantly reducing the memory footprint of Whisper Pro ASR and finalizing the storage cleanup strategy for users migrating from version 1.0.0.
🚀 Key Improvements
🧠 RAM Usage Optimization
- Streamed Ingestion: Completely refactored the file upload pipeline. Media files are now streamed directly from the network to disk, eliminating large RAM spikes previously caused by in-memory buffering.
- Quantized Inference Defaults: Standardized on
int8quantization for CPU and NPU backends, halving the memory required for model weights while maintaining high transcription accuracy. - Proactive Memory Recovery: Integrated explicit garbage collection and hardware cache clearing (CUDA/NPU) into the transcription lifecycle. Memory is now returned to the OS immediately after processing.
- Dynamic UVR Offloading: The heavy vocal isolation engine (UVR/MDX-NET) is now automatically offloaded from memory when idle, freeing up several hundred megabytes of RAM/VRAM for other system tasks.
- ONNX Runtime Tuning: Applied memory-reuse and pattern-matching optimizations to the preprocessing sessions to reduce peak allocation overhead.
🧹 Storage & SSD Endurance
- Extra-Large File Optimization: Implemented a dynamic persistent storage fallback (SSD/HDD) for uploads that exceed
tmpfscapacity. The system now automatically migrates large media ingestion to disk, ensuring stability on low-RAM devices during massive file transfers. - Legacy Cleanup: Version 1.0.2 now automatically detects and purges orphaned 'preprocessing' directories and persistent temp artifacts in the persistent
./model_cachevolume. - Refined Tmpfs Usage: Optimized the interaction between streamed uploads and
tmpfsstorage to ensure maximum SSD protection without risking system stability.
🛠️ Internal Changes
- Version: Bumped to
1.0.2. - Backend: Enhanced
PreprocessingManagerwith model lifecycle controls. - Routes: Optimized
_handle_uploadwith zero-copy stream-to-disk logic.
📦 Deployment
Update your docker-compose.yml to use the 1.0.2 tag (or latest) and restart the service:
docker compose pull
docker compose up -dOptimized for high-concurrency and long-duration media processing.