

The llama.cpp b12489 pre-release, tagged on GitHub at commit 955c4e8 and published via automated release on November 18, 2026 UTC (November 17 Pacific time), addresses a server and sampler failure triggered when structured-output grammar constraints are combined with the ignore_eos flag.
What changed in b12489
According to the release notes, the update centers on pull request #32377, which prevents a crash in sampling when both grammar and ignore_eos are enabled simultaneously. During review, maintainers shifted the conflict check directly into the sampler layer, updated llama-server to reject the combination with an HTTP 400 error, and removed redundant test cases.
The build is distributed as a signed GitHub pre-release accompanied by build attestations and 40 packaged assets spanning CPU, Vulkan, CUDA 12 and 13, ROCm 10.0, OpenVINO, SYCL, Snapdragon Hexagon NPU and Adreno GPU, iOS XCFramework, Android, and s390x targets. Two optional build matrix entries, macOS Apple Silicon with KleidiAI enabled and openEuler, are marked disabled in the release listing.
Why the validation fix matters for deployments
In local and self-hosted inference servers, grammar restricts token generation to match schemas such as JSON, while ignore_eos forces generation past normal end-of-sequence boundaries. Allowing conflicting sampling parameters to reach the sampler without validation can crash shared inference processes, disrupting concurrent clients.
Returning an HTTP 400 response moves the failure to the request boundary rather than terminating the server process. Client applications and agent frameworks that previously sent both options together without noticing an error—or that relied on undocumented behavior—will now receive an explicit bad-request rejection and must disable one of the two flags. Because b12489 is tagged as a pre-release and the release page lists only the change summary and build matrix, operators should validate compatibility in staging before rolling it into production pipelines.
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 2026-11-18.



