Whisper Model & Hardware Acceleration Config

V2sub 2.0 Model Configuration Overview

Navigate to Config → Model Config in the left sidebar to configure compute devices, anti-hallucination parameters, and manage offline Whisper models:

---

Configuration Options Explained

1. Compute Device:
   • GPU Preferred (Default): Automatically uses NVIDIA CUDA acceleration when a compatible GPU is detected, or falls back to CPU.
   • Force CPU: Forces CPU inference even if a GPU is present—useful when your GPU has limited VRAM (e.g., 4GB) but you still want to run the large-v3 model using system RAM.
2. CPU Execution Mode (Pro+):
   • Balanced Mode: Uses ~50% CPU threads so your system remains responsive for other tasks.
   • Full-Speed Mode: Allocates ~90% CPU threads, boosting CPU transcription speed by 30%–50%.
3. Special Optimization (Max+): Tuned Whisper decoding parameters tested across hundreds of challenging videos to prevent repetition during noisy or silent passages.
4. Protected Mode (Max+): Runs inference in isolated anti-hallucination mode. Recommended when a specific video suffers from severe Whisper looping or hallucinations.
5. Model Manager & Open Model Directory: Download any of the 6 models with one click (with instant hot-reload in v2.0.2) or open the local models folder for manual installation.