Add Op(aten::_scaled_dot_product_efficient_attention) | feat(torchlib) #1197
Conversation
titaiwangms
added
the
module: torchlib
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## main #1197 +/- ##
==========================================
- Coverage 78.66% 78.60% -0.07%
==========================================
Files 118 118
Lines 15445 15473 +28
Branches 2428 2431 +3
==========================================
+ Hits 12150 12162 +12
- Misses 2897 2915 +18
+ Partials 398 396 -2 ☔ View full report in Codecov by Sentry. |
|
@justinchuby I tried to run the test on dev10, but it seems the tests are somehow fixed with CPU. Could you point me the direction to test it with GPU? |
|
Maybe try changing ? Apparently we assumed cpu everywhere |
| del kwargs | ||
|
|
||
| make = opinfo_core.partial( | ||
| opinfo_core.make_tensor, device=device, dtype=dtype, requires_grad=requires_grad |
There was a problem hiding this comment.
You may also control device here
There was a problem hiding this comment.
I enabled the whole CUDA tests if it's needed by the test.
| nn_ops.aten__scaled_dot_product_efficient_attention, | ||
| trace_only=True, | ||
| tolerance={torch.float32: (3e-4, 1.5e-5)}, | ||
| # Output[0] is OK, but other outputs just have the same shape with zero values |
There was a problem hiding this comment.
Use the other compare option instead
There was a problem hiding this comment.
compare_shape_only_for_output
|
@justinchuby You might want to review again. I changed the function implementation and enabled CUDA tests. |
|
LGTM. Thanks! |
Fix #1160
Follow up #1043 It's another
scaled_dot_product_XXX_attentionIt only supports CUDA.
https://github.com/pytorch/pytorch/blob/38ae17d166a001ef6837553d1ddffa111624df27/torch/_meta_registrations.py#L5195-L5236
NOTE: This PR also enables CUDA tests.