Segfault possibly caused by changes to MaxOut in v8.1.1
#771
Comments
|
As further info, looking at the traceback again: @RobertSamoilescu has noticed the failure appears to originate in |
Unlikely, using uninitialized memory erroneously wouldn't cause segfaults, only garbage output. We found a bunch of issues in BLIS 0.9 where Thanks for reporting this! |
|
Can you check whether you still see the same issue in an environment that has thinc 0.8.1 + blis 0.7.8? That would let us confirm for sure whether the issue is coming from thinc or blis. |
|
I've run some more CI runs here. The results are as follows:
The error is rather intermittent so not 100% conclusive but this does suggest an issue with more recent |
|
Hi! We've released thinc 8.1.2 today which restricts blis to |
|
Thanks for the update @svlandeg! |
|
@ascillitoe Could you reproduce the crash with the following env variable and post the |
Sure thing, I'll have a go! |
|
Hi again. I've set some more CI runs going here (508 to 518). Interesting, it looks like all the runs that fail (like this one) have: whereas the runs that pass (like this one) have: So it looks like it might be the |
|
Thanks for the quick response! Yeah, the OOB access is in one of the |
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
The blis 0.9.x packages cause segmentation faults: explosion/thinc#771 (comment) For this reason, Thinc/spaCy (the only users of this package in nixpkgs) have reverted to 0.7.x until this issue is resolved.
|
The version constraint for BLIS was updated in efda3b4. Does that mean this issue was fixed? |
|
Thanks for bringing this up. I didn't test to confirm whether this has been fixed when I updated the Blis release to build against numpy v2. I'll have a look at the Blis (upstream) tracker. |
|
Closing this as it should be fixed. Please rereport if the problem comes up again. |
Problem
We have recently (since Sep 9th) been experiencing intermittent seg faults in our (Alibi's) Windows CI. These occur in some of our tests that use the en_core_web_md pipeline.
We believe we have narrowed down the cause to the
Tok2Veccomponent. Since our errors started on the same day v8.1.1 was released we are wondering if the changes toMaxOutare the cause (#702)? Or perhaps the move toblisv0.9 (#736)?We've struggled to come up with a MWE, but for a comparison, we've repeated our CI 30x with v8.1.0 and 30x with v8.1.1. With v8.1.0 we experience no failures, whilst with v8.1.1 we experience seg faults >20% of the time.
Error traceback
From this CI workflow:
Platform, versions etc
windows-2022, version20220905.13.9.12,3.9.13The text was updated successfully, but these errors were encountered: