Non-BLAS works, BLAS crashes. CPU only.

[pegleGrot]

Using the v1.8.3 precompiled binaries, the BLAS variant crashes.
It does show the usual console output but crashes after the "main: processing ..." line:

Details

whisper_init_from_file_with_params_no_state: loading model from 'models\ggml-small-q5_1.bin' whisper_init_with_params_no_state: use gpu = 1 whisper_init_with_params_no_state: flash attn = 1 whisper_init_with_params_no_state: gpu_device = 0 whisper_init_with_params_no_state: dtw = 0 whisper_init_with_params_no_state: devices = 2 whisper_init_with_params_no_state: backends = 2 whisper_model_load: loading model whisper_model_load: n_vocab = 51865 whisper_model_load: n_audio_ctx = 1500 whisper_model_load: n_audio_state = 768 whisper_model_load: n_audio_head = 12 whisper_model_load: n_audio_layer = 12 whisper_model_load: n_text_ctx = 448 whisper_model_load: n_text_state = 768 whisper_model_load: n_text_head = 12 whisper_model_load: n_text_layer = 12 whisper_model_load: n_mels = 80 whisper_model_load: ftype = 9 whisper_model_load: qntvr = 1 whisper_model_load: type = 3 (small) whisper_model_load: adding 1608 extra tokens whisper_model_load: n_langs = 99 whisper_model_load: CPU total size = 189.49 MB whisper_model_load: model size = 189.49 MB whisper_backend_init_gpu: device 0: BLAS (type: 3) whisper_backend_init_gpu: device 1: CPU (type: 0) whisper_backend_init_gpu: no GPU found whisper_backend_init: using BLAS backend whisper_init_state: kv self size = 18.87 MB whisper_init_state: kv cross size = 56.62 MB whisper_init_state: kv pad size = 4.72 MB whisper_init_state: compute buffer (conv) = 22.42 MB whisper_init_state: compute buffer (encode) = 33.85 MB whisper_init_state: compute buffer (cross) = 6.20 MB whisper_init_state: compute buffer (decode) = 97.28 MB

system_info: n_threads = 4 / 4 | WHISPER : COREML = 0 | OPENVINO = 0 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | OPENMP = 1 | REPACK = 1 |

main: processing 'a.flac' (1220376 samples, 76.3 sec), 4 threads, 1 processors, 5 beams + best of 5, lang = he, task = transcribe, timestamps = 1 ...


I tried also using newer OpenBLAS DLLs, v0.3.31, both x64 and x64-64 (no idea what's that supposed to be).

Running on Intel 6th gen (Skylake), no dGPU, Windows 8.1.

Any idea what it might be, or how to analyze further?

[ggerganov]

I can't reproduce on Mac and Linux and don't have Windows setup to investigate. Did it work with any of the earlier versions?

[pegleGrot]

I'm not sure how it was in the past. Will try older versions.
How further back are current model/ggml BINs compatible?

And should older OpenBLAS versions be compatible?
If so, I'll try these too.

[pegleGrot]

BLAS works in versions 1.5 and 1.6, and crashes since 1.7 (tested 1.7.6), which introduced ggml-blas.dll.

I tried libopenblas.dll from v1.5.0 in whisper.cpp v1.8.3. Unrelated to the above crash, it's incompatible:

The ordinal 9544 could not be located in the dynamic link library ggml-blas.dll. 

Tangential observations:

In v1.5 I saw no speed difference between BLAS and non-BLAS.
But very minimal testing, only: ggml-small.bin, one sample audio file, Intel Skylake CPU.
So not sure in which scenarios it's actually useful.

In v1.6 BLAS is substantially slower than non-BLAS.
Maybe libopenblas.dll there is with optimizations off, as the DLL is half the size than in other versions (but still pretty huge).

BTW, the transcription output is somewhat different between the various versions I tried, using the same model file.
I don't know, maybe that's as expected.