Skip to content

The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.

ci(release): whisper's kernels build for AVX, not AVX2, so the editor launches on Ivy Bridge

Jareerb2525370d15c
diff --git a/.github/workflows/build-app.yml b/.github/workflows/build-app.yml
index fc1e5b15..fa64734c 100644
--- a/.github/workflows/build-app.yml
+++ b/.github/workflows/build-app.yml
@@ -73,10 +73,20 @@ env:
# ggml (whisper.cpp's kernels) tunes itself to the machine it is built
# on unless told otherwise, and these bundles run on other machines: a
# runner with AVX-512 or SVE would leave a binary that faults on a CPU
- # without them. Off, ggml compiles to the portable baseline instead -
- # AVX2 with FMA and F16C on x86_64, NEON on ARM. whisper-rs-sys forwards
- # every GGML_* variable to CMake as a define.
+ # without them. Off, ggml does not probe the runner - but its own
+ # instruction-set switches still default to on, which on x86_64 means
+ # AVX2 with FMA, F16C and BMI2: the whole CPU backend is compiled with
+ # /arch:AVX2 (or -mavx2) and the editor dies at launch with an illegal
+ # instruction on a Sandy Bridge or Ivy Bridge machine, which has AVX
+ # and nothing newer (#193). The baseline is AVX: every x86_64 CPU since
+ # 2011 runs it, and whisper on the CPU is somewhat slower for the rest.
+ # NEON on ARM, where these switches are not read. whisper-rs-sys
+ # forwards every GGML_* variable to CMake as a define.
GGML_NATIVE: "OFF"
+ GGML_AVX2: "OFF"
+ GGML_FMA: "OFF"
+ GGML_F16C: "OFF"
+ GGML_BMI2: "OFF"
jobs:
app: