v0.4.1
snarktank/ralphv0.4.1Nov 10, 2025by github-actions[bot]
AI Summary
Fixes a critical memory issue on Blackwell GPUs where the RoPE (Rotary Positional Embedding) cache was incorrectly expanding to an excessive size (~4e18 positions), causing the application to crash.
Key Highlights
- Fixed RoPE cache expansion bug on Blackwell GPUs.
- Resolved memory blow-up issue causing application instability.
New Features
- Bug fix for RoPE cache management.
Full Release Notes
## Bug Fixes - avoid bogus RoPE expansions on Blackwell GPUs https://github.com/TimmyOVO/deepseek-ocr.rs/issues/24 Blackwell builds were blowing up memory because the RoPE cache tried to grow to ~4e18 positions.