
llama.cpp’s b12558 release includes a narrowly targeted Vulkan correctness fix for Qualcomm’s proprietary driver. The upstream report describes incorrect reduction results that can corrupt normalization during decoding and lead to stray generated tokens after long prompts. This is a backend issue; it is not a newly released language model.
The merged change is timestamped 04:58 UTC on November 20. GitHub’s release API records the b12558 publication at 06:38 UTC. Those timestamps distinguish the code merge from the availability of the tagged release.
What changed in the implementation
The code diff adds a driver check that clears the device’s subgroup-arithmetic flag for the Qualcomm proprietary driver. That disables the affected subgroup-based reductions and the related fused addition-and-normalization path. It does not turn off Vulkan for all hardware.
The contributor’s explanation ties the failure to subgroup addition after divergent execution. This is a reminder that a backend can load a model and produce output while still performing part of the calculation incorrectly. Successful startup is therefore a weaker validation than checking the numerical operations and representative generation tasks.
Correctness results are not a general speed claim
The upstream report says selected operation tests on Adreno 850 and X2-90 improved from 20 of 25 to 25 of 25 for ADD_RMS_NORM, and from 33 of 45 to 45 of 45 for RMS_NORM_MUL_ADD. It reports unchanged prompt-processing and generation throughput in one sampled 3B Q4_0 workload. These are contributor results under the stated conditions; OpenSourceCoder.com did not rerun them.
Our analysis: a correctness workaround can be valuable even when it does not increase tokens per second. An application that receives more output is not better served if the output reflects a faulty kernel. At the same time, one unchanged throughput example cannot guarantee the performance of every model, prompt length or Adreno device.
What to check in an affected setup
Developers investigating a matching symptom should record the llama.cpp build, GPU, driver and selected backend before comparing behavior. A CUDA installation on unrelated hardware should not be described as affected merely because it uses the same application. Nor should every odd answer on a Qualcomm machine be assigned to this bug without reproducing the relevant execution path.
The project remains available under its MIT license. Its release page supplies the build matrix and downloads; we did not install a new runtime or modify the shared production model service for this report. Readers can inspect the source through our llama.cpp repository profile.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Original reporting based on primary source review. Publisher results are attributed; no local installation, inference or independent benchmark was performed.



