whisper-stream exits on startup using Vulkan on Windows with AMD 7900 XT

[tokyovigilante]

Hi,

I'm trying to use whisper.cpp with the large-v3 model on an AMD system (7900 XT, Win 10). However after building and starting whisper-stream, the tool exits without any debug logging. Running with -ng works but performance is very poor (just a few random thank-yous). Running under Linux with the same hardware and Vulkan driver works well.

PS C:\Users\ryan\Speech\whisper.cpp> .\build\bin\Release\whisper-stream.exe -m .\ggml-large-v3.bin -t 8  --step 500 --length 5000
init: found 1 capture devices:
init:    - Capture device #0: 'Headset Microphone (Logitech Mono H650e)'
init: attempt to open default capture device ...
init: obtained spec for input device (SDL Id = 2):
init:     - sample rate:       16000
init:     - format:            33056 (required: 33056)
init:     - channels:          1 (required: 1)
init:     - samples per frame: 1024
whisper_init_from_file_with_params_no_state: loading model from '.\ggml-large-v3.bin'
whisper_init_with_params_no_state: use gpu    = 1
whisper_init_with_params_no_state: flash attn = 0
whisper_init_with_params_no_state: gpu_device = 0
whisper_init_with_params_no_state: dtw        = 0
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 7900 XT (AMD proprietary driver) | uma: 0 | fp16: 1 | warp size: 64 | shared memory: 32768 | matrix cores: KHR_coopmat
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       = 51866
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 1280
whisper_model_load: n_audio_head  = 20
whisper_model_load: n_audio_layer = 32
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 1280
whisper_model_load: n_text_head   = 20
whisper_model_load: n_text_layer  = 32
whisper_model_load: n_mels        = 128
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 5 (large v3)
whisper_model_load: adding 1609 extra tokens
whisper_model_load: n_langs       = 100
PS C:\Users\ryan\Speech\whisper.cpp>

[ggerganov]

Does it work with whisper-cli?

[tokyovigilante]

I haven't tried, but was successful getting it to work with hipBLAS with good accuracy.

[ggerganov]

Ok, that rules out a problem with SDL. Maybe after the latest sync (#2868) which includes some Vulkan updates, you can give it another try.

[tokyovigilante]

Thanks, have repulled the latest changes and Vulkan is now working well. Doesn't seem quite as accurate as hipBLAS but haven't tested them head-to-head with recorded audio, and not sure why one would be better than the other. Unfortunately neither as good as my existing proprietary solution for medical dictation but that is the model not the tool. Thanks again for the help!