Crash after "Operation finished" #381

151SoGood · 2025-01-15T03:39:13Z

Whenever I run the large-v3-turbo model, once it completes (as it indicates 'Operation finished in: XX'), cmd prompt hangs & the .exe crashes ('Standalone Faster-Whisper-XXL has stopped working'), and after I click the Windows prompt 'Close program', cmd prompt unfreezes and I can enter the next command.

I don't experience this with the other models, e.g. medium & medium.en.
Any way for us to troubleshoot this? I ran it with '--verbose=true' - anything look odd to you? I can see it glitches at the end, repeating the same text ("Thanks") 6 times. Also, these warnings "Compression ratio threshold is not met with temperature"?

Not a big deal as it does the job to completion, just the crashes are annoying.

Standalone Faster-Whisper-XXL r245.1 running on: CUDA

Number of visible GPU devices: 1

Supported compute types by GPU: {'int8', 'float32', 'int8_float32'}

Note: 'large-v3' model may produce worse results than 'large-v2'!

[2025-01-15 02:47:42.740] [ctranslate2] [thread 53040] [info] CPU: AuthenticAMD (SSE4.1=true, AVX=true, AVX2=true, AVX512=false)
[2025-01-15 02:47:42.740] [ctranslate2] [thread 53040] [info] - Selected ISA: AVX2
[2025-01-15 02:47:42.741] [ctranslate2] [thread 53040] [info] - Use Intel MKL: false
[2025-01-15 02:47:42.741] [ctranslate2] [thread 53040] [info] - SGEMM backend: DNNL (packed: false)
[2025-01-15 02:47:42.741] [ctranslate2] [thread 53040] [info] - GEMM_S16 backend: none (packed: false)
[2025-01-15 02:47:42.742] [ctranslate2] [thread 53040] [info] - GEMM_S8 backend: DNNL (packed: false, u8s8 preferred: true)
[2025-01-15 02:47:42.742] [ctranslate2] [thread 53040] [info] GPU #0: NVIDIA GeForce GTX 1080 (CC=6.1)
[2025-01-15 02:47:42.742] [ctranslate2] [thread 53040] [info] - Allow INT8: true
[2025-01-15 02:47:42.743] [ctranslate2] [thread 53040] [info] - Allow FP16: false (with Tensor Cores: false)
[2025-01-15 02:47:42.743] [ctranslate2] [thread 53040] [info] - Allow BF16: false
[2025-01-15 02:47:45.588] [ctranslate2] [thread 53040] [info] Using CUDA allocator: cuda_malloc_async
[2025-01-15 02:47:45.824] [ctranslate2] [thread 53040] [info] Loaded model C:\Testing\Faster-Whisper-XXL_models\faster-whisper-large-v3-turbo on device cuda:0
[2025-01-15 02:47:45.824] [ctranslate2] [thread 53040] [info] - Binary version: 6
[2025-01-15 02:47:45.825] [ctranslate2] [thread 53040] [info] - Model specification revision: 3
[2025-01-15 02:47:45.826] [ctranslate2] [thread 53040] [info] - Selected compute type: int8_float32

Faster-Whisper's large-v3-turbo model loaded in: 3.15 seconds

Starting sequential inference to transcribe: 'C:\Testing\lorem_ipsumh.mp4'

Processing audio with duration 02:07.640

VAD [silero_v4_fw] finished in: 0:00:00.466

VAD filter removed 00:00.000 of audio

VAD filter kept the following audio segments: [00:00.000 -> 02:07.640]

Processing segment at 00:00.000
[2025-01-15 02:47:46.806] [ctranslate2] [thread 37480] [info] Loaded cuBLAS library version 12.1.3
[00:00.000 --> 00:06.980] lorem ipsum
[00:07.220 --> 00:12.400] So, lorem ipsum
[00:12.620 --> 00:18.720] In terms of where lorem ipsum
[00:19.060 --> 00:23.140] So, lorem ipsum
[00:23.340 --> 00:27.660] And lorem ipsum
Processing segment at 00:27.660
[00:27.660 --> 00:33.920] So, lorem ipsum
[00:34.540 --> 00:40.180] Afterwards, lorem ipsum
[00:40.280 --> 00:43.280] I lorem ipsum
[00:43.500 --> 00:47.860] And lorem ipsum
[00:48.240 --> 00:53.060] And lorem ipsum
[00:53.220 --> 00:55.700] If lorem ipsum
[00:56.140 --> 00:57.360] It'll lorem ipsum.
Processing segment at 00:57.660
[00:57.960 --> 00:59.920] I lorem ipsum
[01:00.140 --> 01:05.080] So lorem ipsum
[01:05.080 --> 01:06.620] before lorem ipsum
[01:06.900 --> 01:09.280] And ilorem ipsum
[01:09.640 --> 01:13.660] lorem ipsum
[01:14.320 --> 01:18.580] lorem ipsum
[01:18.740 --> 01:22.760] lorem ipsum
Processing segment at 01:22.760
[01:22.760 --> 01:28.160] lorem ipsum
[01:28.520 --> 01:33.000] lorem ipsum
[01:33.360 --> 01:36.060] lorem ipsum
Processing segment at 01:36.060
[01:36.060 --> 01:40.660] lorem ipsum
[01:40.960 --> 01:44.260] And lorem ipsum
[01:44.600 --> 01:47.380] And lorem ipsum
Processing segment at 01:47.380
[01:48.720 --> 01:55.320] And lorem ipsum
[01:55.320 --> 01:58.340] to lorem ipsum
[01:58.480 --> 02:00.120] So, lorem ipsum
[02:00.400 --> 02:01.120] This lorem ipsum
[02:01.240 --> 02:04.720] If lorem ipsum
[02:04.740 --> 02:06.820] And lorem ipsum
Processing segment at 02:06.820

    Compression ratio threshold is not met with temperature 0.0 (14.754386 > 1.900000)
    Compression ratio threshold is not met with temperature 0.2 (12.578947 > 1.900000)
    Compression ratio threshold is not met with temperature 0.4 (14.754386 > 1.900000)
    Compression ratio threshold is not met with temperature 0.6 (6.809524 > 1.900000)
    Compression ratio threshold is not met with temperature 0.8 (8.900000 > 1.900000)
    Log probability threshold is not met with temperature 0.8 (-1.429833 < -0.700000)
    Compression ratio threshold is not met with temperature 1.0 (8.669903 > 1.900000)
    Log probability threshold is not met with temperature 1.0 (-0.995539 < -0.700000)
    [02:06.920 --> 02:07.000] Thanks.
    [02:07.000 --> 02:07.100] Thanks.
    [02:07.100 --> 02:07.280] Thanks.
    [02:07.280 --> 02:07.440] Thanks.
    [02:07.440 --> 02:07.620] Thanks.
    Reset prompt. prompt_reset_on_temperature threshold is met 1.000000 > 0.500000

Transcription speed: 6.46 audio seconds/s

Subtitles are written to 'C:\Testing' directory.

Operation finished in: 0:00:23.732

The text was updated successfully, but these errors were encountered:

Purfview · 2025-01-15T05:10:58Z

Whenever I run the large-v3-turbo model, once it completes (as it indicates 'Operation finished in: XX'), cmd prompt hangs & the .exe crashes ('Standalone Faster-Whisper-XXL has stopped working'), and after I click the Windows prompt 'Close program', cmd prompt unfreezes and I can enter the next command.

It's known issue for some users, but reason is unknown and no solution.
It was reported by a user there: #282 (comment).
Then in the newer version I just added "del model" at very end and he said that he didn't reproduced crash anymore, probably that thing was unrelated as there were many libs updated...

You could post your Windows version, GPU model and error from "Windows Event Viewer".

Self-note:
There (#282 (comment)) was this pair of errors from Windows Event Viewer:

Faulting application name: faster-whisper-xxl.exe, version: 192.3.4.0, time stamp: 0x6626da66
Faulting module name: KERNELBASE.dll, version: 10.0.17763.6054, time stamp: 0xc9a93043
Exception code: 0xe06d7363
Fault offset: 0x0000000000041b39

Faulting application name: faster-whisper-xxl.exe, version: 192.3.4.0, time stamp: 0x6626da66
Faulting module name: ucrtbase.dll, version: 10.0.17763.1490, time stamp: 0x48ac8393
Exception code: 0xc0000409
Fault offset: 0x000000000006e77e

Reports at faster-whsiper & ctranslate2:
Windows process crashes when the GPU model is unloaded
Python process crashes on exit under Windows with CUDA

Purfview · 2025-01-15T05:22:05Z

Btw, is that paste from console is same like you see it in console?
Because after Processing segment at 02:06.820 line formatting looks weird, it should look like this:

[00:47.740 --> 00:48.940]  Some text.
  Processing segment at 00:53.480
* Compression ratio threshold is not met with temperature 0.0 (9.054054 > 2.400000)
* Compression ratio threshold is not met with temperature 0.2 (14.565217 > 2.400000)

And there should be two spaces in front of "Processing segment".

Also, these warnings "Compression ratio threshold is not met with temperature"?

That's normal, not warnings, just an info. Don't use verbose if you don't want to see it.

I can see it glitches at the end, repeating the same text ("Thanks") 6 times.

That's "normal" too AKA hallucination. Try better VAD like --vad_method pyannote_v3 it may help.

151SoGood · 2025-01-15T06:47:44Z

Thanks for looking into it! Same pair of errors for me:

Event Viewer logs:

5:41:05PM:
Faulting application name: faster-whisper-xxl.exe, version: 245.1.0.0, time stamp: 0x67859210
Faulting module name: KERNELBASE.dll, version: 10.0.19041.1741, time stamp: 0xe9b4a91b
Exception code: 0xe06d7363
Fault offset: 0x0000000000034fd9

5:41:07PM:
Faulting application name: faster-whisper-xxl.exe, version: 245.1.0.0, time stamp: 0x67859210
Faulting module name: ucrtbase.dll, version: 10.0.19041.789, time stamp: 0x2bd748bf
Exception code: 0xc0000409
Fault offset: 0x000000000007286e

In your 2nd link, there seems to be users who reported a workaround here: SYSTRAN/faster-whisper#71 (comment)

however not sure if that can be implemented as a fix in your version?

151SoGood · 2025-01-15T06:49:26Z

Btw, is that paste from console is same like you see it in console? Because after Processing segment at 02:06.820 line formatting looks weird, it should look like this...

You are right - might have been a copy+paste formatting which messed it up (I think formatting / tab characters got messed up when I copied from my reply in the other issue).

Thanks for also clarifying on the other points, and also providing the suggestion for a better VAD, will give it a go!

Purfview · 2025-01-15T12:24:39Z

Faulting module name: KERNELBASE.dll

Exactly same fault as for another user, it's in Windows' API module, that can mean anything...

To debug further you could share minidump created for a crash:

Select Start > Control Panel.
Double-click System, and then select Advanced system settings > Advanced.
Under Startup and Recovery, select Settings.
In the Write debugging information list, select Small memory dump (256k):

Note: Don't select "Automatically restart".
After a crash minidump should be created at "C:\Windows\Minidump".
Zip it and drag & drop it here into a post.

In your 2nd link, there seems to be users who reported a workaround here: SYSTRAN/faster-whisper#71 (comment)

And users who reported that this workaround doesn't work for them...
One of the workarounds posted there you can try yourself -fallback None, but quality of results may degrade a bit.

Purfview · 2025-01-15T13:25:42Z

Test this "test1" exe: https://fromsmash.com/.x_pTRmmXT-bt
Copy exe to "Faster-Whisper-XXL" folder and check if crash can be reproduced with it.

151SoGood · 2025-01-15T14:07:44Z

Thanks @Purfview. Tested it with test1.exe and it crashed in the same way. Confirmed however there was no minidump created after the crash.

151SoGood · 2025-01-15T14:10:24Z

And users who reported that this workaround doesn't work for them... One of the workarounds posted you can try yourself -fallback None, but quality of results may degrade a bit.

Can confirm this works running -fallback=None avoids the crashing.

Purfview · 2025-01-15T14:10:40Z

Confirmed however there was no minidump created after the crash.

Did you enabled minidump like I wrote? Try to restart Windows.

151SoGood · 2025-01-15T14:11:21Z

Confirmed however there was no minidump created after the crash.

Did you enabled minidump like I wrote? Try to restart Windows.

I did. I will restart and see how that goes.
Update: Will do that later tomorrow when back on the PC.

Purfview · 2025-01-17T18:49:09Z

Closing. Will be fixed when CTranslate2/Nvidia/Windows fix it.

Purfview changed the title ~~Crash using the large-v3-turbo model~~ Crash after transcription is finished Jan 15, 2025

Purfview changed the title ~~Crash after transcription is finished~~ Crash after "Operation finished" Jan 15, 2025

Purfview closed this as completed Jan 17, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Crash after "Operation finished" #381

Crash after "Operation finished" #381

151SoGood commented Jan 15, 2025

Purfview commented Jan 15, 2025

Purfview commented Jan 15, 2025 •

edited

Loading

151SoGood commented Jan 15, 2025 •

edited by Purfview

Loading

151SoGood commented Jan 15, 2025 •

edited by Purfview

Loading

Purfview commented Jan 15, 2025 •

edited

Loading

Purfview commented Jan 15, 2025 •

edited

Loading

151SoGood commented Jan 15, 2025

151SoGood commented Jan 15, 2025 •

edited by Purfview

Loading

Purfview commented Jan 15, 2025

151SoGood commented Jan 15, 2025 •

edited

Loading

Purfview commented Jan 17, 2025

Crash after "Operation finished" #381

Crash after "Operation finished" #381

Comments

151SoGood commented Jan 15, 2025

Purfview commented Jan 15, 2025

Purfview commented Jan 15, 2025 • edited Loading

151SoGood commented Jan 15, 2025 • edited by Purfview Loading

151SoGood commented Jan 15, 2025 • edited by Purfview Loading

Purfview commented Jan 15, 2025 • edited Loading

Purfview commented Jan 15, 2025 • edited Loading

151SoGood commented Jan 15, 2025

151SoGood commented Jan 15, 2025 • edited by Purfview Loading

Purfview commented Jan 15, 2025

151SoGood commented Jan 15, 2025 • edited Loading

Purfview commented Jan 17, 2025

Purfview commented Jan 15, 2025 •

edited

Loading

151SoGood commented Jan 15, 2025 •

edited by Purfview

Loading

151SoGood commented Jan 15, 2025 •

edited by Purfview

Loading

Purfview commented Jan 15, 2025 •

edited

Loading

Purfview commented Jan 15, 2025 •

edited

Loading

151SoGood commented Jan 15, 2025 •

edited by Purfview

Loading

151SoGood commented Jan 15, 2025 •

edited

Loading