fix "Expected all tensors to be on the same device, but found at least two devices" error #11690
There are no files selected for viewing
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
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
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Outdated suggestions cannot be applied.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.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
we shouldn't need that since both hidden_states and res_hidden_states should be on the same device no ? The pre-forward hook added by accelerate should be move all the inputs to the same device.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@SunMarc , i suppose this is a corner case?
torch.catis a weight-less function, so seems cannot covered by the pre-forward hook set byaccelerate...There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I mean since
hidden_statesandres_hidden_states_tupleare in the forward definition, they should be moved to the same device by the pre-forward hook added by accelerateUh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@SunMarc We run into a corner case here. Since we have 8 cards here, so the determined
device_map(by https://github.com/huggingface/diffusers/blob/1bc6f3dc0f21779480db70a4928d14282c0198ed/src/diffusers/models/model_loading_utils.py#L64C5-L64C26) isWe can see
UpBlockis not the atomic module, its submodules are assigned to different devices(up_blocks.0.resnets.0, up_blocks.0.resnets.1), so pre-hook forUpBlockwill not help in this case. And sincetorch.catis not pre-hooked(and cannot since it's a function rather than a module?), so the issue happens.If there is no a
torch.catbtw the sub-blocks inUpBlock, things will be all fine.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@SunMarc, need your inputs in how to proceed for this corner case, thx.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@SunMarc We can see a similar case in
transformersutpytest -rA tests/models/chameleon/test_modeling_chameleon.py::ChameleonVision2SeqModelTest::test_model_parallel_beam_searchw/ 2 cards, the error log is "RuntimeError: Expected all tensors to be on the same device, but found at least two devices,src/transformers/models/chameleon/modeling_chameleon.py", the reason is evenresidualis in the same device ashidden_statesat the beginning, but after they went through some operators as both input and output, they finally placed to different device, but when they come to+which is not a nn.Module(so accelerate cannot pre-hook it), error happens. Do you have some insights on such issues?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@SunMarc, could you share your insights on the issue i mentioned above? thx very much.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sorry for the long wait @yao-matrix , if you add
UpBlock2Din_no_split_modulesofUNet2DConditionModel, the test should pass !_no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"]I've tested on my end and it works.
As for ChameleonVision2SeqModelTest, we probably need to also update _no_split_modules.