No SGEMM backend on CPU despite OpenBLAS installed and CTranslate2 built with -DWITH_OPENBLAS=ON

Summary

I’ve rebuilt CTranslate2 from source (v4.6.0) with -DWITH_OPENBLAS=ON for CPU inference using faster-whisper, but consistently hit this runtime error:

RuntimeError: No SGEMM backend on CPU

Despite verifying that OpenBLAS is installed, SGEMM is available, and everything compiles cleanly.

System Info
OS: Ubuntu 24.04
Python: 3.11.12 (venv)
CTranslate2 version: 4.6.0 (built from source)
BLAS: OpenBLAS installed (libopenblas.so.0)
Model: faster-whisper-base from HuggingFace

Reproduction

1. Built CTranslate2 with:

cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_CUDA=OFF -DWITH_MKL=OFF -DWITH_OPENBLAS=ON
make -j$(nproc)

2. Installed Python bindings:

CMAKE_ARGS="-DCT2_ROOT=$HOME/CTranslate2/build" pip install . --no-build-isolation

3. Confirmed SGEMM exists in OpenBLAS:

import ctypes
blas = ctypes.cdll.LoadLibrary("libopenblas.so.0")
print(blas.cblas_sgemm)  # Found

4. Ran a minimal test script:

from faster_whisper import WhisperModel

print("Loading model...")
model = WhisperModel("/home/andre/faster-whisper-models/models--Systran--faster-whisper-base/snapshots/...", device="cpu", compute_type="float32")
print("Model loaded.")
segments, _ = model.transcribe("debug_compare.wav")

5. Hit this error:

RuntimeError: No SGEMM backend on CPU

Notes

• `ct2-translator` shows the same model path works.
  

• OpenBLAS is installed system-wide and `LD_PRELOAD` had no effect.
  

• `libopenblas.so` is detected in `ldd` of dependent `.so` files.
  

Request
What else can I try to ensure CTranslate2 detects and uses SGEMM from OpenBLAS correctly on CPU?

Thanks in advance.

Reproduce the Ubuntu 24.04 build with the listed CMake options, then compare the working ct2-translator path with the Python binding and faster-whisper script. Check the ctypes SGEMM lookup and ldd results while tracing runtime backend selection; done means the float32 CPU model loads and transcribes without “No SGEMM backend on CPU”.