Closed micmelesse closed 3 months ago
What code does this part of the description:
this pr works around an issue with bias leading to segfaults on some configs.
Might be good to add an assert if the size of the bias tensor is > 2^31.
@vgokhale Bad wording on my part. I have updated the description and the assert.
The core dump mentioned in the removed skip was from the issue fixed here right? Do we still need to run a separate UT over parameterizing bias?
@groenenboomj There is a ut with bias. see here. The github diff view is not great because the test is similar to the previous test.
@groenenboomj
@micmelesse please correct me if I'm wrong, but this isn't a "core dump issue" as such. The issue here is that triton cannot address a tensor that is > 2^31 elements. Note that it is not that Triton cannot address > 2^31 bytes. We can have several tensors in memory of size 2^30, and this will be fine. But a single tensor cannot be > 2^31.