[dynamic shapes] stop writing Max(*, 1) for strides #150376
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/150376
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (2 Unrelated Failures)As of commit 3d7ac71 with merge base 9810488 ( BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
…torch into pianpwk/max_1_strides
| @@ -885,7 +885,7 @@ def test_non_overlapping_and_dense_unbacked(self): | |||
| cf( | |||
| torch.empty_strided( | |||
| (2, 3, 1, u0), | |||
| (3 * Max(1, u0), Max(1, u0), Max(1, u0), 1), | |||
| (3 * u0, u0, u0, 1), | |||
There was a problem hiding this comment.
this test case is no longer relevant
| sym_shape_meta.strides_[last_idx] = c10::SymInt(1); | ||
| for (auto i = last_idx - 1; i >= 0; --i) { | ||
| sym_shape_meta.strides_[i] = sym_shape_meta.strides_[i + 1] * | ||
| sym_shape_meta.sizes_[i + 1].max(1); | ||
| if (TORCH_GUARD_OR_TRUE(sym_shape_meta.sizes_[i + 1].sym_gt(1))) { |
There was a problem hiding this comment.
we used to generate max(u0, 1)
and now we generate just u0 unconditionally?
can you explain why this is safe?
There was a problem hiding this comment.
Back to your queue to address @laithsakka's comments around safety
@laithsakka @pianpwk I wonder for these guard_or_true cases if we should by default always add a comment about safety just like how rust does this with unsafe blocks: https://std-dev-guide.rust-lang.org/policy/safety-comments.html
When handling strides, avoiding generating Max(u0, 1) expressions if we can, since it'll be hard to deal with these once we move away from guard_size_oblivious.
Looking at what the code was before sym_max was introduced (#94400 in
_prims_common/__init__.py), this change seems appropriate.