Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Inference & Quantization / Developer & Agent Tools

llama.cpp b12498 Fixes Server LoRA Scheduling and Hybrid SWA Tests

Pre-release b12498 of llama.cpp addresses LoRA and recurrent state handling in the server slot scheduler and aligns hybrid SWA checkpoint tests to run one token per ubatch.

Listen to this articleGoogle Voice
0:00
GitHub release social card for ggml-org/llama.cpp b12498
AI-generated conceptual illustration by AiBrainWorX; not a photograph or product screenshot.
Event / context: 2026-11-17. Source check: October 28, 2026.

The ggml-org maintainers have published llama.cpp release b12498 as a pre-release build (commit 5297534), focusing on server-side slot scheduling reliability for LoRA adapters and recurrent architectures alongside a targeted test fix for hybrid Sliding Window Attention (SWA) checkpoints.

Slot Scheduler Fix for LoRA and Recurrent Models

In pull request #32434, the update modifies server: slot_sched to resolve bugs affecting LoRA adapters and recurrent model states in llama-server, paired with corresponding updates to the server LoRA and recurrent test suites. For deployments serving multiple requests or mixing adapter and recurrent workloads through the built-in HTTP server, slot scheduling determines how context slots and state caches are assigned and reused across active sequences.

Why Hybrid SWA Checkpoint Tests Needed Exact Ubatch Parity

Release b12498 also adjusts the hybrid SWA checkpoint test to process one token per micro-batch (ubatch). According to the commit notes, cached and uncached runs previously split prompts into different ubatches and processed generated tokens differently—decoding one token at a time when cached versus batching them when uncached.

The maintainers noted that on CUDA, this execution-path difference shifted top logprobs on a small test model by up to 0.03 and swapped near-tie tokens, triggering false test failures even though the cache itself was functioning properly. Enforcing one token per ubatch ensures both test paths invoke the same kernels so outputs can be compared directly.

Platform Targets and Availability

Because b12498 is tagged as an automated GitHub Actions pre-release rather than a major version milestone, teams running stable production builds should evaluate whether the slot scheduler fix affects their server workloads before upgrading. The release matrix includes prebuilt artifacts and attestations across CPU, Vulkan, CUDA 12 and 13, ROCm 10.0, SYCL, OpenVINO, and Snapdragon backends on Linux, Windows, macOS, iOS, and Android, while the macOS Apple Silicon KleidiAI build variant is marked disabled in this build.

EXPLORE THE SOURCE

GitHub repositories

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-18.

Follow your curiosity.

Pass the signal.