ggml-org/llama.cpp is a project that performs large language model inference in C/C++. The release b12034, published on November 2, 2026, fixed an issue in the calculation of WebGPU temporary memory allocation size. According to the release summary, it fixed the problem where the size was not computed monotonically with respect to batch length, preventing consecutive allocation errors. The core of this update is not the addition of new inference features but the correction of that memory size calculation. This description is sourced from the b12034 release page on GitHub.
The target of the fix is the calculation of temporary memory allocation size used in WebGPU. The problem was explicitly stated to be that the size calculation depending on batch length was not monotonic, which could lead to consecutive allocation errors. b12034 is the change that fixes that calculation problem and prevents the errors. Therefore, what can be confirmed in this release is limited to the specific allocation problem on the WebGPU path and its fix. Whether other execution paths or separate feature changes are included in this release cannot be verified from the provided release summary alone.
The provided information does not state the specific conditions under which the error occurred or the scope of impact. No performance changes or benchmark results are presented either, so there is no basis to evaluate the effect of the fix numerically. What can be confirmed is the release description stating that it fixed the non-monotonic temporary memory size calculation depending on batch length and prevented consecutive allocation errors. Therefore, it is appropriate to view b12034 as a fix for WebGPU allocation errors, and no further impact or performance improvement can be concluded from this change alone.