Stop force realizing to prevent recursion errors unless it's much bigger #138881
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/138881
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit a1f0c78 with merge base 8fbf866 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| def test_inductor_no_recursionerror_on_for_loops(self): | ||
| def forward(x): | ||
| for _ in range(1000): | ||
| for _ in range(10000): |
There was a problem hiding this comment.
This already wasn't failing without the PR intended to address this. I presume some other thing ameliorated the issue, so I bumped up the threshold to make sure we would hit the recursion error unless we explicitly break it up.
|
|
||
| mask_mod = mk_3d_flex_natten_mask(data_size, kernel_size) | ||
|
|
||
| torch.compile(create_block_mask)(mask_mod, None, None, S, S) |
There was a problem hiding this comment.
This compile create_block_mask a side effect?
There was a problem hiding this comment.
no, it's mainly just validating that it runs correctly in Inductor haha
|
|
||
| def has_large_inner_fn(self): | ||
| return self.inner_fn_opcount().num_ops > config.realize_opcount_threshold | ||
| def has_large_inner_fn(self, threshold=None): |
There was a problem hiding this comment.
| def has_large_inner_fn(self, threshold=None): | |
| def has_large_inner_fn(self, threshold : Optional[int] = None) -> bool: |
There was a problem hiding this comment.
Nit: why not set default of threshold to be 0 in the first place since you need set it to 0 if it's None.
|
|
||
| def has_large_inner_fn(self): | ||
| return self.inner_fn_opcount().num_ops > config.realize_opcount_threshold | ||
| def has_large_inner_fn(self, threshold=None): |
There was a problem hiding this comment.
Nit: why not set default of threshold to be 0 in the first place since you need set it to 0 if it's None.
|
@shunting314 I think it's unlikely across our dashboard, since it would require a string of unbroken pointwise ops that's long enough. |
4 participants
Stack from ghstack (oldest at bottom):
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov @rec