Geoff Tognetti a9bfb9c771 Fix model storage location and Parakeet CUDA errors ...

This commit includes two related fixes:

1. Model Storage Location Fix:
   - Models were stored outside persistent volume, causing re-downloads on every container rebuild
   - Implemented dependency injection for adapter constructors
   - Adapters now accept envPath parameter instead of hardcoded paths
   - Added registerAdapters() function in main.go to register adapters after config loads
   - Removed init() functions from all adapters (whisperx, parakeet, canary, sortformer, pyannote)
   - Models now persist in /app/data/whisperx-env (inside Docker volume)

2. Parakeet CUDA Error 35 Fix:
   - Fixed CUDA_ERROR_ILLEGAL_ADDRESS on long audio files (30+ minutes)
   - Disabled CUDA graphs via change_decoding_strategy() in buffered inference
   - Fixed buffered script check in PrepareEnvironment() to verify both transcribe.py and transcribe_buffered.py exist
   - Verified working: 49min audio transcribed in 63sec (47x realtime)

Files Modified:
- cmd/server/main.go: Added registerAdapters() function
- internal/transcription/adapters/whisperx_adapter.go: Constructor accepts envPath, removed init()
- internal/transcription/adapters/parakeet_adapter.go: Constructor accepts envPath, removed init(), added CUDA fix and buffered script check
- internal/transcription/adapters/canary_adapter.go: Constructor accepts envPath, removed init()
- internal/transcription/adapters/sortformer_adapter.go: Constructor accepts envPath, removed init()
- internal/transcription/adapters/pyannote_adapter.go: Constructor accepts envPath, removed init()

2025-11-15 14:00:35 -08:00

..

main.go

Fix model storage location and Parakeet CUDA errors

2025-11-15 14:00:35 -08:00