Fix the bug of using loss before assignment #700
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Outdated suggestions cannot be applied.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.
Bug Description
A bug is triggered when using MoE in conjunction with pipeline_parallel_size > 1, resulting in a 'referenced before assignment' error.
The complete error report is as follows
Script
The above BUG can be reproduced using my script examples/pretrain_gpt_moe_demo.sh
Solution
Upon examining the error code, I noticed a potential issue in line 216 of megatron/core/pipeline_parallel/schedules.py:
config.grad_scale_func(torch.tensor(1.0, device=loss.device))
The variable loss runs the risk of being referenced before assignment. Therefore, I suggest modifying it as per my pull request:
config.grad_scale_func(torch.tensor(1.0, device=output_tensor.device))