Ban tensor constants in tracing. #1844
Open
Conversation
This file contains 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
Summary:
There're two patterns resulting in the similar constant lifting effect during tracing:
```
def forward(self, x):
return x + torch.tensor(...)
```
and
```
class Model(torch.nn.Module):
def __init__(self):
super().__init__()
self.a = torch.tensor(...)
def forward(self, x):
return x + self.a
```
In theory and practice, it will be too implicit to promote them as buffers in tracing, due to two reasons:
1. They don't have a formal FQN.
2. They likely result in baked-in values without being noticed by users.
3. Their values can be easily mutated without being noticed by users.
Therefore we want to make a policy decision to ban any form of tensor constants to show up in the traced graph. To proceed in the tracing, users either need to register these constants explicitly as buffers, or use constructors to properly initialize them. In some cases, users can get away without even wrapping scalars inside a torch.tensor() call.
Differential Revision: D53436364
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D53436364 |
facebook-github-bot
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Feb 6, 2024
Summary: X-link: pytorch/executorch#1844 There're two patterns resulting in the similar constant lifting effect during tracing: ``` def forward(self, x): return x + torch.tensor(...) ``` and ``` class Model(torch.nn.Module): def __init__(self): super().__init__() self.a = torch.tensor(...) def forward(self, x): return x + self.a ``` In theory and practice, it will be too implicit to promote them as buffers in tracing, due to two reasons: 1. They don't have a formal FQN. 2. They likely result in baked-in values without being noticed by users. 3. Their values can be easily mutated without being noticed by users. Therefore we want to make a policy decision to ban any form of tensor constants to show up in the traced graph. To proceed in the tracing, users either need to register these constants explicitly as buffers, or use constructors to properly initialize them. In some cases, users can get away without even wrapping scalars inside a torch.tensor() call. Test Plan: buck test mode/opt caffe2/test:test_export -- -r test_export_tensor_constants Differential Revision: D53436364
facebook-github-bot
pushed a commit
to pytorch/pytorch
that referenced
this pull request
Feb 6, 2024
Summary: X-link: pytorch/executorch#1844 There're two patterns resulting in the similar constant lifting effect during tracing: ``` def forward(self, x): return x + torch.tensor(...) ``` and ``` class Model(torch.nn.Module): def __init__(self): super().__init__() self.a = torch.tensor(...) def forward(self, x): return x + self.a ``` In theory and practice, it will be too implicit to promote them as buffers in tracing, due to two reasons: 1. They don't have a formal FQN. 2. They likely result in baked-in values without being noticed by users. 3. Their values can be easily mutated without being noticed by users. Therefore we want to make a policy decision to ban any form of tensor constants to show up in the traced graph. To proceed in the tracing, users either need to register these constants explicitly as buffers, or use constructors to properly initialize them. In some cases, users can get away without even wrapping scalars inside a torch.tensor() call. Test Plan: buck test mode/opt caffe2/test:test_export -- -r test_export_tensor_constants Differential Revision: D53436364
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.
Summary:
There're two patterns resulting in the similar constant lifting effect during tracing:
and
In theory and practice, it will be too implicit to promote them as buffers in tracing, due to two reasons:
Therefore we want to make a policy decision to ban any form of tensor constants to show up in the traced graph. To proceed in the tracing, users either need to register these constants explicitly as buffers, or use constructors to properly initialize them. In some cases, users can get away without even wrapping scalars inside a torch.tensor() call.
Differential Revision: D53436364