llama.cpp b12441 Pre-release Fixes WebGPU Mask Bug
Posted As of
llama.cpp pre-release b12441 (2026-11-16) fixed a bug where the WebGPU flash-attention shader read the mask without its head dimension. b12436, released the same day, fixed a Vulkan 32-bit compilation issue.
What happened
Source: ggml-org/llama.cpp GitHub releases page
- Pre-release b12441 was tagged on 2026-11-16 (official)
- The WebGPU flash-attention shader read the mask without a head dimension, which the release notes say produced incorrect results when the mask had two or more heads
- The fix adds the head index to the mask offset
- Includes a change that makes Vulkan top-k tie selection deterministic. The release notes explain that the radix select using an atomic counter could leave different columns depending on thread scheduling
- b12436 (2026-11-16) fixed a Vulkan 32-bit compilation issue
Our take
For users running local models on PC through the Vulkan or WebGPU paths, this fix concerns attention result accuracy and result reproducibility. However, the list of affected models, performance figures, and how much output differed in earlier builds have not been confirmed. Any direct impact on Mac or glasses devices is not confirmed in the notes, and the individual tag pages were not opened.
Sources and related posts
- ggml-org/llama.cpp GitHub releases Primary source