Copy of the inputs for the evaluation during compilation #1150
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
|
||||||||||||||
|
||||||||||||||
vkovinicTT
force-pushed
the
vkovinic/eval_tensor_copy
branch
from
363f22e to
3413ebf
Compare
|
||||||||||||||
|
||||||||||||||
vkovinicTT
force-pushed
the
vkovinic/eval_tensor_copy
branch
3 times, most recently
from
94ab912 to
4cf2184
Compare
vkovinicTT
changed the title
vkovinicTT
requested review from
mstojkovicTT,
nvukobratTT,
pilkicTT and
dgolubovicTT
as code owners
|
||||||||||||||
|
||||||||||||||
vkovinicTT
force-pushed
the
vkovinic/eval_tensor_copy
branch
from
4cf2184 to
9565677
Compare
|
||||||||||||||
1 similar comment
|
||||||||||||||
|
||||||||||||||
1 similar comment
|
||||||||||||||
nvukobratTT
reviewed
Feb 19, 2025
| return x + y | ||
|
|
||
| input = torch.zeros(shape, requires_grad=False) | ||
| framework_input = input.detach().clone() |
There was a problem hiding this comment.
Why we're doing detach and clone here?
There was a problem hiding this comment.
Because if run inference on the torch module and then want to run it on compiled module, inputs for the compiled module would be different due to the changes made by torch (as default torch is not functional, meaning it will change input)
| compiled_model = forge.compile(framework_model, sample_inputs=tt_inputs, module_name="inplace") | ||
| tty = compiled_model(*tt_inputs)[0] | ||
|
|
||
| compare_with_golden(golden=y, calculated=tty) |
There was a problem hiding this comment.
Should we use the standard verify function?
There was a problem hiding this comment.
For the nature of the in-place problem we are facing, I have separate input for compiled module and torch module so I can't pass it to the verify function.
| tty = compiled_model(*tt_inputs)[0] | ||
|
|
||
| compare_with_golden(golden=y, calculated=tty) | ||
| print(framework_input) |
There was a problem hiding this comment.
Do we need these prints?
| # convert tensor from tf to torch | ||
| y = torch.tensor(y.numpy()) | ||
|
|
||
| compare_with_golden(golden=y, calculated=tty) |
There was a problem hiding this comment.
Similar comments as for PT example
|
I would still wait with this PR until we resolve which policy we want to follow (functional or per framework policy). |
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
This PR ensures that a copy of the input tensors is created when running forward during compilation (e.g. when trying to verify outputs using forward pass of the framework model). This prevents unintended modifications to the original input tensors if in-place operations are performed during the forward pass.
In order for it to work, first this PR in the
tt-tvmneeds to be merged.Why is this needed?
Some models perform in-place operations on input tensors during the forward pass, which can lead to unintended changes in the original inputs. By making a copy of the inputs, we ensure correctness and avoid potential issues when running forward during compilation.
Example test:
❗❗❗ IMPORTANT NOTES ❗❗❗
In the training mode if the inputs require grad, pytorch will try to do the forward and will throw runtime error:
a leaf Variable that requires grad is being used in an in-place operation., while the compiled model silently computes the gradients and this need to be addressed.Introducing this change our compiler will not change it's input, so it will act as tensorflow (it doesn't change it's input tensor), but it is won't be aligned with pytorch which allows in-place changes of the input tensor.