Make Transformer tolerate missing layers for PP #322
Merged
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
facebook-github-bot
added
the
CLA Signed
wconstab
added a commit
that referenced
this pull request
May 10, 2024
A few small changes here lets manual PP frontend 'reconfigure' a whole transformer model to a stage's portion simply by setting undesired layers to None (in cases of top level layers) or deleting them from the ModuleDict (for 'layers.*'). These changes don't impact the FQNs of the remaining layers, which is critical for checkpoint load/save compatibility. ghstack-source-id: 48a7aaf Pull Request resolved: #322
tianyu-l
reviewed
May 11, 2024
| h = self.tok_embeddings(tokens) if self.tok_embeddings else tokens | ||
|
|
||
| for layer in self.layers: | ||
| for layer in self.layers.values(): |
Contributor
There was a problem hiding this comment.
Is order still respected after switching to dict? If not, we need to sort the layers based on int(key).
Contributor
Author
There was a problem hiding this comment.
Contributor
|
Nice. But it is less intuitive than I originally thought. Especially the int/str conversion part. Not sure if that's a best UX for pippy or a customized PipelineModuleList will be easier for users. |
fegin
approved these changes
May 13, 2024
wconstab
added a commit
that referenced
this pull request
May 13, 2024
A few small changes here lets manual PP frontend 'reconfigure' a whole transformer model to a stage's portion simply by setting undesired layers to None (in cases of top level layers) or deleting them from the ModuleDict (for 'layers.*'). These changes don't impact the FQNs of the remaining layers, which is critical for checkpoint load/save compatibility. ghstack-source-id: 48a7aaf Pull Request resolved: #322
awgu
reviewed
May 14, 2024
| self.layers = torch.nn.ModuleDict() | ||
| for layer_id in range(model_args.n_layers): | ||
| self.layers.append(TransformerBlock(layer_id, model_args)) | ||
| self.layers[str(layer_id)] = TransformerBlock(layer_id, model_args) |
Collaborator
There was a problem hiding this comment.
curious why do the dict keys have to be str (as opposed to int directly)?
Collaborator
|
One downside to using |
tianyu-l
pushed a commit
to tianyu-l/torchtitan_intern24
that referenced
this pull request
Aug 16, 2024
A few small changes here lets manual PP frontend 'reconfigure' a whole transformer model to a stage's portion simply by setting undesired layers to None (in cases of top level layers) or deleting them from the ModuleDict (for 'layers.*'). These changes don't impact the FQNs of the remaining layers, which is critical for checkpoint load/save compatibility. ghstack-source-id: 48a7aaf Pull Request resolved: pytorch#322
philippguevorguian
pushed a commit
to YerevaNN/YNNtitan
that referenced
this pull request
Aug 17, 2024
A few small changes here lets manual PP frontend 'reconfigure' a whole transformer model to a stage's portion simply by setting undesired layers to None (in cases of top level layers) or deleting them from the ModuleDict (for 'layers.*'). These changes don't impact the FQNs of the remaining layers, which is critical for checkpoint load/save compatibility. ghstack-source-id: 48a7aaf Pull Request resolved: pytorch#322
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.
Stack from ghstack (oldest at bottom):
A few small changes here lets manual PP frontend 'reconfigure' a whole
transformer model to a stage's portion simply by setting undesired
layers to None (in cases of top level layers) or deleting them from the
ModuleDict (for 'layers.*').
These changes don't impact the FQNs of the remaining layers, which is
critical for checkpoint load/save compatibility.