[Relay][Op] Trilu operator implementation #12124
Conversation
|
@sfvaroglu Can you take a look at this PR? |
|
LGTM! Thanks for doing this @jwfromm! cc @mikepapadim |
jwfromm
force-pushed
the
trilu_op
branch
2 times, most recently
from
221026d to
cc2864e
Compare
| "test_tril_neg", | ||
| "test_tril_one_row_neg", | ||
| "test_tril_out_neg", | ||
| "test_tril_out_pos", | ||
| "test_tril_zero", |
There was a problem hiding this comment.
I haven't looked at this op at all. How tricky would it be to support the zero case? Otherwise LGTM.
There was a problem hiding this comment.
It actually works on llvm and cuda. I was testing on my macbook and it seems like the metal backend in general doesnt support empty tensors. I think for CI we could add these cases.
There was a problem hiding this comment.
Seems like it also doesnt work with nvptx for the same issue with empty tensors. I'll add them here and see how it does in CI.
|
I added pytorch testing and integration. Thanks for the recommendation @shingjan. |
|
Unfortunately the empty tensor tests still fail on CI GPUs. I'm not sure why, it doesnt seem like its directly related to this PR so I'm going to reenable skips for those tests. |
|
@mbrookhart I think this is ready to merge. |
|
Thanks @jwfromm @sfvaroglu @mikepapadim @shingjan |
* Added topi trilu implementation * Implemented and tested full Trilu op. * Fix test type. * Add tril zero tests. * Add pytorch trilu integration. * Clean up torch integration. * Readded skip for zero tests.
This PR adds a new operator that supports triangular masking similar to that in
np.triuandnp.tril. The addition ofrelay.triluconveniently lets us pass many of the remaining onnx tests.