[Quant][ONEDNN] Fix weight reorder issue for grouped convolution #91934
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/91934
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 4 PendingAs of commit f2ab2a5: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
github-actions
bot
added
the
module: cpu
Xia-Weiwen
force-pushed
the
onednn_grouped_conv_reorder
branch
from
0b46451 to
b8f546d
Compare
|
Shall we target this PR to PT 2.0 release? |
OK |
Xia-Weiwen
requested review from
digantdesai,
jerryzh168,
jianyuh,
kimishpatel and
salilsdesai
as code owners
Xia-Weiwen
force-pushed
the
onednn_grouped_conv_reorder
branch
2 times, most recently
from
bec298b to
1ac56bd
Compare
Xia-Weiwen
force-pushed
the
onednn_grouped_conv_reorder
branch
from
1ac56bd to
6c35c60
Compare
|
@pytorchbot merge |
Merge failedReason: Approval needed from one of the following (Rule 'superuser'): Details for Dev Infra teamRaised by workflow job |
|
Hi @jerryzh168. Could you review? Thanks. |
|
Hi @jerryzh168. Could you please take a look of this PR? It involves onednn backend only. Thanks! |
…onednn quant backend
Xia-Weiwen
force-pushed
the
onednn_grouped_conv_reorder
branch
from
6c35c60 to
f2ab2a5
Compare
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
5 participants
Summary
For onednn quant backend only.
QConv weight may be reordered to another blocked format if input shape is changed at runtime. It's a bug that group info is not retained for such reordering. This may lead to wrong shape of weight after reordering. This PR fixes this bug.
Test plan
python test/test_quantization.py -k test_conv_reorder_issue_onednn
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10