ggerganov/whisper.cpp

Crash if --processors > 8

Closed this issue · 5 comments

jrosener commented

I am writing an application that is able to transcribe multiple audio in parallel using the same model.
For that I use one common whisper_context for multiple whisper_state used by worker threads where transcriptions processing are performed with whisper_full_with_state().
It works perfectly until 8 parallel transcriptions but crashes into whisper_full_with_state() if running more transcriptions.

Because this implementation is based on whisper_full_parallel() used by the main sample application, it is possible to reproduce the issue by running it using more than 8 --processors: ./build/bin/main --model ggml-tiny.bin --processors 9 10min_audio_french.wav
It does not matter if it is running on cpu, openvino, cuda,...: it always crashes.

Results:

$ ./build/bin/main --model ggml-tiny.bin --processors 9 10min_audio_french.wav 
whisper_init_from_file_with_params_no_state: loading model from 'ggml-tiny.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
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 384
whisper_model_load: n_audio_head  = 6
whisper_model_load: n_audio_layer = 4
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 384
whisper_model_load: n_text_head   = 6
whisper_model_load: n_text_layer  = 4
whisper_model_load: n_mels        = 80
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 1 (tiny)
whisper_model_load: adding 1608 extra tokens
whisper_model_load: n_langs       = 99
whisper_model_load:      CPU total size =    77.11 MB
whisper_model_load: model size    =   77.11 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB

system_info: n_threads = 36 / 8 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | METAL = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | CUDA = 0 | COREML = 0 | OPENVINO = 0 | CANN = 0

main: processing '10min_audio_french.wav' (9600000 samples, 600.0 sec), 4 threads, 9 processors, 5 beams + best of 5, lang = en, task = transcribe, timestamps = 1 ...

whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
whisper_init_state: kv self size  =    3.15 MB
whisper_init_state: kv cross size =    9.44 MB
whisper_init_state: kv pad  size  =    2.36 MB
whisper_init_state: compute buffer (conv)   =   13.19 MB
whisper_init_state: compute buffer (encode) =   64.79 MB
whisper_init_state: compute buffer (cross)  =    3.88 MB
whisper_init_state: compute buffer (decode) =   95.89 MB
Segmentation fault (core dumped)

Questions:

  • Are 8 parallel transcriptions (so 8 running whisper_state for 1 common context) a known limitation ?
  • If yes, is there a way to increase that limitation ?
  • Or is it a bug ?
jrosener commented

Is #2509 the same issue ? (probably)

KitaitiMakoto commented

When I increase GGML_MAX_CONTEXTS in ggml/include/ggml.h, main doesn't crash. 8 is magic number but there should be some limitation.

refs: #2520

jrosener commented

When I increase GGML_MAX_CONTEXTS in ggml/include/ggml.h, main doesn't crash. 8 is magic number but there should be some limitation.

refs: #2520

Indeed, it works well by increasing this value (tested at 256, so running 32 parallel transcriptions). It would be great to make it possible to change it in whisper_context_params or somewhere else.

ggerganov commented

With #2525 merged there is no longer limit on the context, so it should work with any number of processors.

  • 🎉1
jrosener commented

Yes, no limitation anymore. Many thanks !