A user has submitted a fix for llama.cpp that resolves crashes and stalls encountered when running DeepSeek V4 Flash (DSV4). The issue stemmed from improper sequence removal during divergent turns, which caused the system to re-prefill thousands of tokens instead of just the delta.
- The root cause was DSV4's inability to perform partial seq_rm, forcing a checkpoint restore that failed to purge future SWA cells.
- This error exhausted the 768-cell SWA window, leading to a "Context size has been exceeded" crash around logical depth 1250 tokens.
- The fix implements a proper seq_rm of the diverged suffix, ensuring only the delta is re-prefilled.
- The author provides a forked branch addressing these issues for reproduction and testing.
The update removes both the crash and the stall, improving user experience even at lower token speeds.