b10442

ggml-org/llama.cppb10442Aug 15, 2026by github-actions[bot]same day as b10441

What changed since b10441

low risk

Vulkan backend gains Intel Xe shared-memory and coopmat1 reshape optimizations, with no CLI or API changes.

New features

  • Vulkan coopmat1 matmul now supports SHMEM_STRIDE_PAD and APPLY_SLM_A_RESHAPE paths for Intel Xe hardware.
  • Shared-memory sizing logic accounts for Intel's SHMEM_STRIDE_PAD=0 case when deciding matmul_shmem_support.

Improvements

  • Shared kvalues_mxfp4 buffers are cacheline-aligned to improve Vulkan memory access efficiency.
  • SLM-A reshape is scoped to the Intel Windows driver only, avoiding unintended use on other Vulkan drivers.
  • Vulkan shared memory estimation corrected for Intel configurations, reducing wasted allocation.

Bug fixes

  • Fixed an out-of-bounds read during kvalues_mxfp4 initialization introduced by the cacheline padding change.
  • Reverted the mxfp4 cacheline padding where it caused problems, keeping only the safe alignment changes.

Upgrade notes

This is a drop-in patch release with no CLI, config, or API changes; the --load-mode migration from b10441 remains in effect. Only Vulkan users, particularly on Intel Xe GPUs, will see behavioral differences, and the SLM-A reshape path activates solely under the Intel Windows driver.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Enhances Vulkan backend performance for Intel Xe hardware with optimized shared memory operations and reshape functionality.

Key Highlights

  • Added SHMEM_STRIDE_PAD support for Intel Xe
  • Added APPLY_SLM_A_RESHAPE for coopmat1
  • Fixed shared memory estimation
  • Cacheline alignment improvements
  • Bug fixes for out-of-bounds reads

Release Notes (excerpt)

<details open> vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat1 on Intel Xe (#25380) * vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat mul_mm on Intel Xe * vulkan: fix shmem estimate for Intel SHMEM_STRIDE_PAD=0 in matmul_shmem_support * cacheline aligned for shared kvalues_mxfp4 * vulkan: fix OOB read in kvalues_mxfp4 init after cacheline padding * vulkan: restrict SLM-A reshape to Intel Windows driver, revert mxfp4 cacheline padding </details> **Website:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories