Change8

b9911

📦 llama-cppView on GitHub →
4 features🐛 3 fixes🔧 4 symbols

Summary

This release focuses heavily on CUDA performance optimizations by fusing MMVQ post-scale operations for NVFP4 and adding dense MMVQ fusion. Several cleanups and reordering steps were performed to stabilize and restrict fusion benefits to appropriate kernels.

Migration Steps

  1. Remove the environment variable previously used to disable fusion, as fusion is now managed differently/enabled by default under certain conditions.

✨ New Features

  • Fused MMVQ post-scale operation for NVFP4 kernels in CUDA.
  • Added dense MMVQ fusion.
  • Enabled fusion for MoE models in shared MMVQ operations.
  • Merged old mul_mat patterns into the lane-based approach.

🐛 Bug Fixes

  • Reordered bias/scale operations to adhere to requirements specified in #24331.
  • Restricted scale_view_nodes and enrolled MM + ADD into lane-matcher.
  • Restricted scale-fusion specifically to NVFP4 kernels due to potential performance regressions on other quantizations caused by a heavy prologue in GEMV.

Affected Symbols