b10704
📦 llama-cppView on GitHub →
✨ 1 features🔧 1 symbols
Summary
Introduced an optimized CUDA path for mm_ids_helper, improving performance for specific expert configurations. This update expands support for more n_expert_used values on the fast path.
✨ New Features
- Optimized CUDA path for mm_ids_helper when n_expert_used is a divisor of warp_size, now supporting n_expert_used = 10 and other previously unsupported counts.