[JIT] Ensure offset is a multiple of 4 to fix "Philox" RNG in jitted kernels#50169
Closed
mcarilli wants to merge 3 commits intopytorch:masterfrom
Closed
[JIT] Ensure offset is a multiple of 4 to fix "Philox" RNG in jitted kernels#50169mcarilli wants to merge 3 commits intopytorch:masterfrom
mcarilli wants to merge 3 commits intopytorch:masterfrom
Conversation
Contributor
💊 CI failures summary and remediationsAs of commit 25cdf69 (more details on the Dr. CI page):
This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions to the (internal) Dr. CI Users group. This comment has been revised 6 times. |
Codecov Report
@@ Coverage Diff @@
## master #50169 +/- ##
==========================================
+ Coverage 80.49% 80.68% +0.18%
==========================================
Files 1900 1900
Lines 206254 206254
==========================================
+ Hits 166018 166409 +391
+ Misses 40236 39845 -391 |
ngimel
approved these changes
Jan 8, 2021
Collaborator
|
Please add the PR description as a note somewhere in the code. |
added 2 commits
January 8, 2021 15:55
ngimel
approved these changes
Jan 9, 2021
Contributor
facebook-github-bot
left a comment
There was a problem hiding this comment.
@ngimel has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Contributor
laurentdupin
pushed a commit
to laurentdupin/pytorch
that referenced
this pull request
Apr 24, 2026
…kernels (pytorch#50169) Summary: Immediately-upstreamable part of pytorch#50148. This PR fixes what I'm fairly sure is a subtle bug with custom `Philox` class usage in jitted kernels. `Philox` [constructors in kernels](https://github.com/pytorch/pytorch/blob/30206b504ed5e786ad2792061ec5ebe4b9b6abe9/torch/csrc/jit/codegen/cuda/codegen.cpp#L102) take the cuda rng generator's current offset. The Philox constructor then carries out [`offset/4`](https://github.com/pytorch/pytorch/blob/677f0d6383cde8700c41a6ca8e69a6f1d9748b4e/torch/csrc/jit/codegen/cuda/runtime/random_numbers.cu#L13) (a uint64_t division) to compute its internal offset in its virtual Philox bitstream of 128-bit chunks. In other words, it assumes the incoming offset is a multiple of 4. But (in current code) that's not guaranteed. For example, the increments used by [these eager kernels](https://github.com/pytorch/pytorch/blob/677f0d6383cde8700c41a6ca8e69a6f1d9748b4e/aten/src/ATen/native/cuda/Distributions.cu#L171-L216) could easily make offset not divisible by 4. I figured the easiest fix was to round all incoming increments up to the nearest multiple of 4 in CUDAGeneratorImpl itself. Another option would be to round the current offset up to the next multiple of 4 at the jit point of use. But that would be a jit-specific offset jump, so jit rng kernels wouldn't have a prayer of being bitwise accurate with eager rng kernels that used non-multiple-of-4 offsets. Restricting the offset to multiples of 4 for everyone at least gives jit rng the chance to match eager rng. (Of course, there are still many other ways the numerics could diverge, like if a jit kernel launches a different number of threads than an eager kernel, or assigns threads to data elements differently.) Pull Request resolved: pytorch#50169 Reviewed By: mruberry Differential Revision: D25857934 Pulled By: ngimel fbshipit-source-id: 43a75e2d0c8565651b0f12a5694c744fd86ece99
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Immediately-upstreamable part of #50148.
This PR fixes what I'm fairly sure is a subtle bug with custom
Philoxclass usage in jitted kernels.Philoxconstructors in kernels take the cuda rng generator's current offset. The Philox constructor then carries outoffset/4(a uint64_t division) to compute its internal offset in its virtual Philox bitstream of 128-bit chunks. In other words, it assumes the incoming offset is a multiple of 4. But (in current code) that's not guaranteed. For example, the increments used by these eager kernels could easily make offset not divisible by 4.I figured the easiest fix was to round all incoming increments up to the nearest multiple of 4 in CUDAGeneratorImpl itself.
Another option would be to round the current offset up to the next multiple of 4 at the jit point of use. But that would be a jit-specific offset jump, so jit rng kernels wouldn't have a prayer of being bitwise accurate with eager rng kernels that used non-multiple-of-4 offsets. Restricting the offset to multiples of 4 for everyone at least gives jit rng the chance to match eager rng. (Of course, there are still many other ways the numerics could diverge, like if a jit kernel launches a different number of threads than an eager kernel, or assigns threads to data elements differently.)