llama.cpp build b10286 ships one grammar change and nearly 30 platform targets. The notable code shift: max repetition values of 2000 or above now degrade to unbounded.
What changed:
- Grammar max repetition (#26613): Setting max repetition to ≥2000 now silently becomes unbounded instead of enforcing the literal cap. Affects grammar-constrained generation where you limit token repetition.
- CUDA 13 support: Windows x64 builds ship with CUDA 13.3 DLLs alongside the existing CUDA 12.4 variant.
- KleidiAI on macOS arm64: DISABLED in this build.
- openEuler base: DISABLED, but Ascend NPU builds (310p, 910b with ACL Graph) remain active for x86 and aarch64.
The grammar change is the one developers need to act on. If you use llama.cpp’s grammar engine to constrain output — JSON schemas, regex, structured generation — and set a high max repetition value as a safety net, that net is gone above 2000. The threshold silently switches to “no limit” rather than enforcing 2000 repetitions.
NOTE
PR #26613 uses the word “degrade” — the behavior intentionally falls back to unbounded when the value is impractically high. If your grammar relies on a specific max repetition cap between 2000 and infinity, verify your output still conforms.
Build b10286 is live across nearly 30 platform targets. The grammar change lands here — check your repetition settings if you use grammar-constrained generation.