← News
Runtime

llama.cpp fixes Metal private-buffer ranges on discrete-GPU Macs and opt-in paths

Merged November 19 and shipped in b12548, #32602 fixes the length argument to NSMakeRange on the private-buffer path. It is not a general Apple Silicon speedup; upstream says this path is outside CI.

Source: llama.cpp GitHub PR / release

llama.cpp merged Metal fix #32602 on November 19 and included it in b12548 during this reporting window. The issue is not parameter count or quantization: the second argument of NSMakeRange is a length, not an end position. Code compiling successfully does not prove that it describes the intended memory range.

  1. Establish whether the affected path is used. The contributor identifies private buffers, affecting discrete GPUs on Intel Macs or environments with GGML_METAL_SHARED_BUFFERS_DISABLE=1. A Metal fix does not imply that every Apple Silicon deployment has the same defect. Record the device, build and environment variables together.

  2. Merge and release are new events, not a new all-device validation campaign. The PR states that these paths are not covered by CI. AraoAI has not independently reproduced the issue, and no matched performance comparison is available here. We report no speedup percentage or quality improvement and do not equate a merged patch with validation of every model and system combination.

  3. Validate the target workload after upgrading. Keep the old binary and launch configuration, then compare load logs and task outputs using the same weights, quantization, context and backend settings. Include non-default environment variables in the reproduction. Do not move an experimental private-buffer configuration into production merely to exercise this fix.

This is a narrowly scoped range-argument correction, not a general performance promise. Engine and model licenses remain separate. Future independent reports should retain the OS version, build, trigger and original failure before broadening the impact assessment.

Source: https://github.com/ggml-org/llama.cpp/pull/32602 Release: https://github.com/ggml-org/llama.cpp/releases/tag/b12548