Enabling oneDNN and OpenBLAS backends at the same time breaks downstream applications
Open
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 4.7k
- Forks
- 536
- Avg merge
- 12h 12m
- Merged PRs (30d)
- 4
Description
When oneDNN and OpenBLAS are enabled the wyoming-faster-whisper does not start and works only with OpenBLAS enabled: https://github.com/NixOS/nixpkgs/pull/237788#issuecomment-1592070847
On the other hand, the libretranslate also breaks when oneDNN and OpenBLAS are enabled and works only with oneDNN enabled:
[2023-06-15 09:09:43.116] [ctranslate2] [thread 535123] [warning] The compute type inferred from the saved model is int8, but the target device or backend do not support efficient int8 computation. The model weights have been automatically converted to use the float32 compute type instead.
[1] 535079 segmentation fault (core dumped)
Here's a little table:
| x86_64-linux | only oneDNN | only OpenBLAS | both oneDNN & oneDNN |
|---|---|---|---|
| libretranslate | works | segmentation fault | segmentation fault |
| wyoming-faster-whisper | RuntimeError | works | RuntimeError |
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Reproduce the failures with oneDNN and OpenBLAS enabled using the libretranslate and wyoming-faster-whisper cases described in the report. Trace backend initialization and inference behavior for the combined configuration; done means both applications start and run without the reported segmentation fault or RuntimeError.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 30/100