ggml
ggml copied to clipboard
ggml-org
ggml_vulkan crashes if running 2 threads which allocate their own context for the same GPU device
I am using Whisper.net which in turn uses whisper.cpp which in turn uses ggml.
Running a single thread with the Vulkan backend works perfectly (but does not saturate the GPU - unlike when using the CUDA backend, which is a separate issue), so I tried adding a second thread which does everything for the same GPU that the first thread does (allocate context and push a separate audio stream). This crashes (invalid memory access), and the 2 thread call stacks always look like this:
Thread
Not Flagged 4488 50 Worker Thread SpeechToTextJob.serial: 1/2 msvcp140d.dll!mtx_do_lock
[External Code]
ggml-vulkan-whisper.dll!ggml_pipeline_allocate_descriptor_sets(std::shared_ptr<vk_device_struct> & device={...}) Line 909
ggml-vulkan-whisper.dll!ggml_backend_vk_graph_compute(ggml_backend * backend=0x0000032c7918ccd0, ggml_cgraph * cgraph=0x0000032bb8c1a138) Line 7589
ggml-base-whisper.dll!ggml_backend_graph_compute_async(ggml_backend * backend=0x0000032c7918ccd0, ggml_cgraph * cgraph=0x0000032bb8c1a138) Line 333
ggml-base-whisper.dll!ggml_backend_sched_compute_splits(ggml_backend_sched * sched=0x0000032c72339bb0) Line 1397
ggml-base-whisper.dll!ggml_backend_sched_graph_compute_async(ggml_backend_sched * sched=0x0000032c72339bb0, ggml_cgraph * graph=0x0000032c7a1a0060) Line 1589
ggml-base-whisper.dll!ggml_backend_sched_graph_compute(ggml_backend_sched * sched=0x0000032c72339bb0, ggml_cgraph * graph=0x0000032c7a1a0060) Line 1572
whisper.dll!ggml_graph_compute_helper(ggml_backend_sched * sched=0x0000032c72339bb0, ggml_cgraph * graph=0x0000032c7a1a0060, int n_threads=4) Line 183
whisper.dll!whisper_encode_internal(whisper_context & wctx={...}, whisper_state & wstate={...}, const int mel_offset=0, const int n_threads=4, bool(*)(void *) abort_callback=0x00007ff87a5f3318, void * abort_callback_data=0x0000000000000001) Line 2350
whisper.dll!whisper_full_with_state(whisper_context * ctx=0x0000032bb8ad0980, whisper_state * state=0x0000032c72c95d70, whisper_full_params params={...}, const float * samples=0x000002eb0af9fa28, int n_samples=462400) Line 5573
[External Code]
Whisper.net.dll!Whisper.net.WhisperProcessor.Process(System.ReadOnlySpan<float> samples = "System.ReadOnlySpan<Single>[462400]", System.DateTime firstSampleTime = "2020-10-01 10:00:28.300", bool audioWasSkipped = false) Line 192
Whisper.net.dll!Whisper.net.WhisperProcessor.Process(float[] samples = {float[462400]}, System.DateTime firstSampleTime = "2020-10-01 10:00:28.300", bool audioWasSkipped = false) Line 160
[External Code]
Not Flagged > 36004 51 Worker Thread SpeechToTextJob.serial: 2/2 ggml-vulkan-whisper.dll!std::string::size
ggml-vulkan-whisper.dll!std::string::size() Line 2346
ggml-vulkan-whisper.dll!std::hash<std::string>::_Do_hash(const std::string & _Keyval) Line 3354
ggml-vulkan-whisper.dll!std::_Conditionally_enabled_hash<std::string,1>::operator()(const std::string & _Keyval) Line 2339
ggml-vulkan-whisper.dll!std::_Uhash_compare<std::string,std::hash<std::string>,std::equal_to<std::string>>::operator()<std::string>(const std::string & _Keyval) Line 151
ggml-vulkan-whisper.dll!std::unordered_map<std::string,std::weak_ptr<vk_pipeline_struct>,std::hash<std::string>,std::equal_to<std::string>,std::allocator<std::pair<std::string const ,std::weak_ptr<vk_pipeline_struct>>>>::at(const std::string & _Keyval) Line 437
ggml-vulkan-whisper.dll!ggml_pipeline_allocate_descriptor_sets(std::shared_ptr<vk_device_struct> & device={...}) Line 912
ggml-vulkan-whisper.dll!ggml_backend_vk_graph_compute(ggml_backend * backend=0x0000032c791b0250, ggml_cgraph * cgraph=0x0000032bb8c20d38) Line 7589
ggml-base-whisper.dll!ggml_backend_graph_compute_async(ggml_backend * backend=0x0000032c791b0250, ggml_cgraph * cgraph=0x0000032bb8c20d38) Line 333
ggml-base-whisper.dll!ggml_backend_sched_compute_splits(ggml_backend_sched * sched=0x0000032bb871d490) Line 1397
ggml-base-whisper.dll!ggml_backend_sched_graph_compute_async(ggml_backend_sched * sched=0x0000032bb871d490, ggml_cgraph * graph=0x0000032cbd850060) Line 1589
ggml-base-whisper.dll!ggml_backend_sched_graph_compute(ggml_backend_sched * sched=0x0000032bb871d490, ggml_cgraph * graph=0x0000032cbd850060) Line 1572
whisper.dll!ggml_graph_compute_helper(ggml_backend_sched * sched=0x0000032bb871d490, ggml_cgraph * graph=0x0000032cbd850060, int n_threads=4) Line 183
whisper.dll!whisper_encode_internal(whisper_context & wctx={...}, whisper_state & wstate={...}, const int mel_offset=0, const int n_threads=4, bool(*)(void *) abort_callback=0x00007ff87a5f3318, void * abort_callback_data=0x0000000000000002) Line 2311
whisper.dll!whisper_full_with_state(whisper_context * ctx=0x0000032bb8ad0980, whisper_state * state=0x0000032bac961410, whisper_full_params params={...}, const float * samples=0x000002eb0b93b5e0, int n_samples=256000) Line 5573
[External Code]
Whisper.net.dll!Whisper.net.WhisperProcessor.Process(System.ReadOnlySpan<float> samples = "System.ReadOnlySpan<Single>[256000]", System.DateTime firstSampleTime = "2020-10-01 10:13:00.000", bool audioWasSkipped = false) Line 192
Whisper.net.dll!Whisper.net.WhisperProcessor.Process(float[] samples = {float[256000]}, System.DateTime firstSampleTime = "2020-10-01 10:13:00.000", bool audioWasSkipped = false) Line 160
[External Code]
Running multiple concurrent threads+contexts for Cuda or CPU or even Vulkan CPU works as expected. Only the Vulkan GPU backend causes the crash. If I add a lock to prevent the 2 threads from calling WhisperProcessor.Process() at the same time, the crash goes away. This happens with the 1.7.3 release from a few weeks ago as well as whisper.cpp+ggml I pulled and built yesterday.
Yeah, this is a known issue (https://github.com/ggerganov/llama.cpp/issues/7575 https://github.com/ggerganov/llama.cpp/issues/11371) that has not yet been fixed.
