propagate .meta info when replacing subgraphs in fx #87255
Closed
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/87255
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 3385d45: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
github-actions
bot
requested review from
Chillee,
Krovatkin,
albanD,
anjali411,
antoniojkim,
ezyang,
miladm and
wconstab
3 tasks
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
wconstab
approved these changes
Oct 19, 2022
Contributor
wconstab
left a comment
wconstab
left a comment
There was a problem hiding this comment.
seems like a good change, but i wonder how we can tell if we're still missing cases, should the default be true?
Contributor
Author
|
Yeah - I was pretty torn about whether or not copying metadata should be explicit vs. implicit. Right now it's explicit - you specify I wasn't sure about making it implicit, because I guess I could flip it so that we copy metadata by default, but only if there's no metadata field set on the new node (and if there is, maybe we should just silently not copy metadata instead of erroring). I'm open to either / other opinions though. |
bdhirsh
commented
Oct 21, 2022
| f'args = {_format_arg(self.args)}, kwargs = {_format_arg(self.kwargs)})' | ||
|
|
||
| @compatibility(is_backward_compatible=True) | ||
| def replace_all_uses_with(self, |
Contributor
Author
There was a problem hiding this comment.
this is an FX public API change (it shouldn't be BC breaking though) - so I'll have to update the tests.
Who else should review public FX API changes? @SherlockNoMad ?
Contributor
There was a problem hiding this comment.
This is not a BC breaking change. So it should be fine.
SherlockNoMad
approved these changes
Oct 21, 2022
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
This was referenced Oct 24, 2022
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
This was referenced Nov 1, 2022
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. [ghstack-poisoned]
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Nov 5, 2022
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. Pull Request resolved: pytorch#87255 Approved by: https://github.com/wconstab, https://github.com/SherlockNoMad
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Dec 10, 2022
Fixes pytorch/torchdynamo#1708 Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new `call_module` node in the graph. If the original subgraph outputs had fake tensors and other metadata stored in their `.meta` attribute though, then this information was getting lost when we spliced in the subgraph. Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the `fx.Node.replace_all_uses_with`, just because propagating metadata seems like a pretty core thing. Pull Request resolved: pytorch#87255 Approved by: https://github.com/wconstab, https://github.com/SherlockNoMad
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Fixes pytorch/torchdynamo#1708
Our FX subgraph partitioner works by taking all of the original output nodes from a subgraph, and replacing it with a new
call_modulenode in the graph.If the original subgraph outputs had fake tensors and other metadata stored in their
.metaattribute though, then this information was getting lost when we spliced in the subgraph.Losing metadata on an FX graph also seems like an easy trap to fall into, so I'm wondering if there are any better guardrails that we can add. I ended up fixing in this PR by adding an optional kwarg to propagate meta info directly in the
fx.Node.replace_all_uses_with, just because propagating metadata seems like a pretty core thing.Stack from ghstack (oldest at bottom):