[DONT MERGE YET][dynamo][guards] Do not dict pop in framelocals dict creation #164038
Conversation
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
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/164038
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 044eaff with merge base 55840fb ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
anijain2305
added
topic: not user facing
Lucaskabela
approved these changes
Sep 29, 2025
…ion"
For this program
```
import torch
torch._dynamo.config.enable_cpp_symbolic_shape_guards = False
def fn(x):
for _ in range(20000):
x = x + 1
return x
mod = torch.fx.symbolic_trace(fn)
print(mod)
opt_mod = torch.compile(mod, backend="eager", dynamic=True)
x = torch.randn(2, 2)
opt_mod(x)
```
It reduces the guard overhead from 1.8 ms to 172 us. The dict pop is not required as explained in the comment.
It can reduce to 46 us if we just remove the _dict.attr pop branch.
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela
[ghstack-poisoned]
williamwen42
approved these changes
Sep 30, 2025
…ion"
For this program
```
import torch
torch._dynamo.config.enable_cpp_symbolic_shape_guards = False
def fn(x):
for _ in range(20000):
x = x + 1
return x
mod = torch.fx.symbolic_trace(fn)
print(mod)
opt_mod = torch.compile(mod, backend="eager", dynamic=True)
x = torch.randn(2, 2)
opt_mod(x)
```
It reduces the guard overhead from 1.8 ms to 72 us. The dict pop is not required as explained in the comment.
It can reduce to 46 us if we just remove the _dict.attr pop branch.
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela
[ghstack-poisoned]
| } | ||
| } else { | ||
| _dict[framelocals_names[i]] = value; | ||
| seen.insert(name_ptr); |
There was a problem hiding this comment.
Actually, is there a guarantee that entries in framelocals_names with the same string value are the string object? If not, then the unordered_set should use Python string compare for its hash/equality? (We should also try to craft a test case that fails if this update is done incorrectly - I recall encountering failing tests when I made mistakes in the initial implementation, so I'm confused as to why the tests today don't seem to catch incorrect updates.)
There was a problem hiding this comment.
Yeah, that was my concern as well. I don't know how to make that test case though.
|
Not merging yet. @williamwen42 has concerns and he will take over this PR. |
anijain2305
changed the title
williamwen42
added a commit
that referenced
this pull request
Sep 30, 2025
Followup to #164038 cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 1, 2025
Followup to #164038 Pull Request resolved: #164316 Approved by: https://github.com/anijain2305
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
) Followup to pytorch#164038 Pull Request resolved: pytorch#164316 Approved by: https://github.com/anijain2305
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Stack from ghstack (oldest at bottom):
For this program
It reduces the guard overhead from 1.8 ms to 72 us. The dict pop is not required as explained in the comment.
It can reduce to 46 us if we just remove the _dict.attr pop branch.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @Lucaskabela