

The llama.cpp b12456 pre-release, tagged on November 16, 2026 (November 17 UTC) at commit d261e49, focuses on stability and correctness in the project's OpenCL backend. The release integrates pull request #32110, which resolves three distinct defects in the bin flash-attention prefill path used by compatible OpenCL devices.
Memory Leak in Multi-Row Prefill and Speculative Decoding
According to the release notes, the WMM flash-attention kernel previously allocated a fallback mask image on every invocation over a K-cache sub-buffer without releasing it afterward. Because each unreleased image kept its underlying K-cache sub-buffer alive, OpenCL objects accumulated over time. Every call processing more than one query row—including every verification step during speculative decoding—followed this path, causing server performance to degrade with each successive request. The update ensures the fallback image is properly released.
Numerical Safeguard for q8_0 K/V at Head Size 512
A second fix keeps q8_0 quantized key-value caches at head dimension DK=512 off the WMM flash-attention kernel. In the release notes, the maintainers report that the bin flash-attention kernel produced incorrect outputs whenever V held different values from K at that shape, citing a maximum error of 2.04 in test-backend-ops FLASH_ATTN_EXT hsk=512 q8_0 on X2-90 hardware, whereas f16 K/V at DK=512 remained accurate. Because the backend's built-in flash-attention kernels already decline q8_0 at DK=512, the change makes behavior consistent regardless of whether the external kernel library is active.
Fallback Beyond Hardware 3D Image Width Limits
The third patch prevents assertion failures on longer context windows. Because the WMM prefill path repacks Q, K, and V tensors into OpenCL 3D images with V spanning n_kv in width, prompts exceeding a GPU's maximum 3D image width—described in the notes as roughly 16k tokens of context on X2-90—caused clCreateImage to fail and trigger an assertion. Shapes past that device limit now fall back cleanly to the regular flash-attention path.
Scope and Availability
Listed as a signed GitHub pre-release with build attestations, b12456 includes prebuilt artifacts across Linux, macOS, Android, and Windows—including Windows arm64 OpenCL Adreno targets—while macOS Apple Silicon KleidiAI and openEuler builds are marked disabled in this release listing. While the fixes do not alter CUDA, Vulkan, or CPU inference paths, they remove key reliability blockers for long-running OpenCL servers, quantized KV caches at large head sizes, and longer-context prefill workloads.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-17.



