[original author: mrnikwaws] fixing incorrect stride information on xla tensors #5468
Conversation
aws-kingrj
commented
Aug 18, 2023
- fixing incorrect xla tensor stride information
- breaks log softmax lowering
- tested on AWS Neuron internal testing
- original author mrnikwaws
|
Ok I think I know what's going on. The pr from the fork can not use remote cache so build just timeout. Let me see if I can bump the timeout. |
|
pr in #5470. Let me also grant you write access so you can create branch on pytorch/xla directly. This way all of the CI will use the bazel remote cache and runs much faster. |
|
@aws-kingrj if you rebase the build should start run pass timeout. |
| @@ -161,6 +161,11 @@ at::IntArrayRef XLATensorImpl::strides_custom() const { | |||
| return strides_default(); | |||
| } | |||
|
|
|||
There was a problem hiding this comment.
You mentioned this breaks log softmax lowering. Can you add a test that will fail without this change? The change in general make sense but without a test it is hard to make sure this is correct nor can we prevent this from regressing again.
| @@ -3700,6 +3700,19 @@ TEST_F(AtenXlaTensorTest, TestPowIntExponent) { | |||
| ExpectCounterNotChanged("aten::.*", cpp_test::GetIgnoredCounters()); | |||
| } | |||
|
|
|||
| TEST_F(AtenXlaTensorTest, TestPowFloatScalarBaseIntExponent) { | |||
There was a problem hiding this comment.
I think the commit is a bit messed up, this seems to be from other pr