Change8

b10241

📦 llama-cppView on GitHub →
2 features🐛 1 fixes🔧 1 symbols

Summary

This release addresses data-races in CUDA block_reduce operations by improving synchronization when reusing shared memory. It also introduces double-buffering optimizations for softmax and norm operations.

✨ New Features

  • Implemented double-buffering for single-row softmax.
  • Implemented double-buffering for norm.

🐛 Bug Fixes

  • Fixed data-races when reusing SMEM in block_reduce by ensuring proper synchronization after reading from SMEM.

Affected Symbols