Skip to content

fix: portable ggml on x86_64 release legs (GGML_NATIVE=OFF) - #5

Merged
ericmann merged 1 commit into
mainfrom
fix/portable-ggml
Oct 29, 2026
Merged

ericmann merged 1 commit into
mainfrom
fix/portable-ggml

Conversation

@ericmann

Copy link
Copy Markdown
Contributor

ext-whisper v0.1.1's x86_64 release binaries were built with -march=native on AVX-512 runners and crash with SIGILL on AVX2-only CPUs (verified on a Ryzen 5 3400G; v0.1.0 had 0 AVX-512 instructions, v0.1.1 has 4,460). whisper-rs-sys passes GGML_* env vars to CMake, so the managed release workflow now exports GGML_NATIVE=OFF on x86_64 legs: ggml's portable AVX2 baseline, matching the platform's x86-64-v3 floor. v0.1.1 was never published; ext-whisper v0.1.2 ships the fix. Ships as ext-template v0.2.1.

whisper.cpp's CMake defaults to GGML_NATIVE=ON, so ext-whisper's vendored
ggml was compiled with -march=native for the release runner. The
ubuntu-26.04 runners have AVX-512: ext-whisper v0.1.1's x86_64 binary
contains 4,460 zmm instructions (v0.1.0: none) and dies with SIGILL on a
Ryzen 5 3400G (AVX2, no AVX-512). whisper-rs-sys forwards GGML_* env vars
to CMake; OFF selects ggml's portable AVX/AVX2/FMA defaults. x86_64 legs
only; arm64 macOS native is the M1 floor. llama-cpp-sys-2 already forces
OFF (ext-infer 0.2.1's binary: 0 zmm, runs on AVX2).
@ericmann
ericmann merged commit fb900b9 into main Oct 29, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant