]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
metal: handle command buffer failures gracefully in synchronize (#20306)
authorJulian Pscheid <redacted>
Tue, 10 Mar 2026 06:32:24 +0000 (23:32 -0700)
committerGitHub <redacted>
Tue, 10 Mar 2026 06:32:24 +0000 (08:32 +0200)
commit1a5631beaa2bfb9f76f177a8dbe5544dd6c2d528
treea45276c070c97a00451ea93dba20bead67b1f125
parent1dab5f5a443a7b972005c56fb92eca2b07d57fea
metal: handle command buffer failures gracefully in synchronize (#20306)

Replace GGML_ABORT("fatal error") in ggml_metal_synchronize() with
error flag + return. This aligns synchronize error handling with
graph_compute, which already returns GGML_STATUS_FAILED for the same
condition.

When a command buffer fails (e.g., iOS GPU access revocation during
backgrounding, macOS eGPU disconnect, OOM), the backend enters an
error state instead of killing the host process. Subsequent
graph_compute calls return GGML_STATUS_FAILED immediately. Recovery
requires recreating the backend.

Failed extra command buffers are properly released on the error path
to avoid Metal object leaks.
ggml/src/ggml-metal/ggml-metal-context.m