[fx] Add offload codegen #1598
Merged
[fx] Add offload codegen #1598
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
Merge ColossalAI
Daily merge
|
Paste an example of the generated code here. |
FrankLeeeee
reviewed
Sep 14, 2022
| if node.name == "linear4": | ||
| setattr(node, "activation_checkpoint", [0]) | ||
|
|
||
| gm = ColoGraphModule(copy.deepcopy(model), graph) |
There was a problem hiding this comment.
I think copy is not needed.
There was a problem hiding this comment.
Copy is needed because we want to test backward gradients.
Here is an example for the generated code: def pack_hook(self, x):
if getattr(x, "offload", None):
return (x.device, x.cpu())
else:
return x
def unpack_hook(self, packed):
if isinstance(packed, tuple):
device, tensor = packed
return tensor.to(device)
else:
return packed
def checkpoint_0(self, linear2):
linear3 = self.linear3(linear2); linear2 = None
linear4 = self.linear4(linear3); linear3 = None
return linear4
def forward(self, x):
linear1 = self.linear1(x); x = None
setattr(linear1, 'offload', True)
with torch.autograd.graph.saved_tensors_hooks(self.pack_hook, self.unpack_hook):
linear2 = self.linear2(linear1); linear1 = None
linear4 = colossalai.utils.activation_checkpoint.checkpoint(self.checkpoint_0, True, linear2, use_reentrant=False)
linear5 = self.linear5(linear4); linear4 = None
return linear5 |
FrankLeeeee
reviewed
Sep 14, 2022
Comment on lines
153
to
154
| @pytest.mark.skipif(with_codegen, reason='torch version is equal to or higher than 1.12.0') | ||
| @pytest.mark.skip(reason="currently torch11 ColoGraphModule is not implemented") |
There was a problem hiding this comment.
Sure I could remove the first one~
added 2 commits
September 14, 2022 14:42
…ossalAI into feature/add_offload_codegen
FrankLeeeee
approved these changes
Sep 14, 2022
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.
What's New
Previously we have offload option in activation checkpoint region to support the offload input process, however, in the upcoming activation solver, we might offload the input of a node that is not inside any checkpoint region. Therefore, I use saved_tensors_hooks for this kind of offload manner.
As we haven't implemented the torch11 version of
ColoGraphModule, I skip the unit test for this part and attach the results on torch12 below