Fix: only add parameter with grads to parameter group#7869
Merged
Conversation
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f0265ef538
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Collaborator
Author
Collaborator
Author
A new test is added, this test will fail on master with the exact error described, and pass with this fix. |
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Collaborator
Author
|
Hi @PKUWZP This PR is ready for review again. A test is added to expose the issue it intend to fix. |
delock
changed the title
tjruwase
approved these changes
Mar 1, 2026
nathon-lee
pushed a commit
to nathon-lee/DeepSpeed_woo
that referenced
this pull request
Mar 7, 2026
This PR fix a bug when Muon optimizer is used on training part of the model parameters. When train part of the model parameters (and freeze all others). In certain case, all trainable paramters will use Muon optimizer and non of them use AdamW optimizer, or vice versa. It will cause one of `muon_params` and `non_muon_params` to contain only non-trainable parameters, which would eventurally cause the following failure. A reasonable fix is only add parameter with grads to `muon_params` and `non_muon_params`, so the case above would cause one of the parameter groups to be empty and get filterd out immediately. ``` [rank3]: Traceback (most recent call last): [rank3]: File "/home/gma/transfer_qwen/finetune_moe.py", line 904, in <module> [rank3]: main(args) [rank3]: File "/home/gma/transfer_qwen/finetune_moe.py", line 709, in main [rank3]: model_engine, optimizer, train_dataloader, lr_scheduler = deepspeed.initialize( [rank3]: ^^^^^^^^^^^^^^^^^^^^^ [rank3]: File "/home/gma/DeepSpeed/deepspeed/__init__.py", line 214, in initialize [rank3]: engine = DeepSpeedEngine(args=args, [rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank3]: File "/home/gma/DeepSpeed/deepspeed/runtime/engine.py", line 363, in __init__ [rank3]: self._configure_optimizer(optimizer, model_parameters) [rank3]: File "/home/gma/DeepSpeed/deepspeed/runtime/engine.py", line 1585, in _configure_optimizer [rank3]: self.optimizer = self._configure_zero_optimizer(basic_optimizer) [rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank3]: File "/home/gma/DeepSpeed/deepspeed/runtime/engine.py", line 1893, in _configure_zero_optimizer [rank3]: optimizer = Stage1And2ZeroOptimizer( [rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^ [rank3]: File "/home/gma/DeepSpeed/deepspeed/runtime/zero/stage_1_and_2.py", line 403, in __init__ [rank3]: flattened_buffer = self.flatten_dense_tensors_aligned( [rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank3]: File "/home/gma/DeepSpeed/deepspeed/runtime/zero/stage_1_and_2.py", line 1044, in flatten_dense_tensors_aligned [rank3]: return self.flatten(align_dense_tensors(tensor_list, alignment)) [rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank3]: File "/raid/gma/miniforge3/envs/ds/lib/python3.12/site-packages/torch/_utils.py", line 571, in _flatten_dense_tensors [rank3]: return torch._C._nn.flatten_dense_tensors(tensors) [rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank3]: ValueError: torch.cat(): expected a non-empty list of Tensors ``` --------- Signed-off-by: Ma, Guokai <guokai.ma@intel.com> Signed-off-by: nathon-lee <leejianwoo@gmail.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
This PR fix a bug when Muon optimizer is used on training part of the model parameters.
When train part of the model parameters (and freeze all others). In certain case, all trainable paramters will use Muon optimizer and non of them use AdamW optimizer, or vice versa. It will cause one of
muon_paramsandnon_muon_paramsto contain only non-trainable parameters, which would eventurally cause the following failure.A reasonable fix is only add parameter with grads to
muon_paramsandnon_muon_params, so the case above would cause one of the parameter groups to be empty and get filterd out immediately.