Add state_dict converter for DeepSeekv3 in torchtitan #1538
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
meta-cla
bot
added
the
CLA Signed
wwwjn
commented
Aug 7, 2025
wwwjn
changed the title
wwwjn
commented
Aug 11, 2025
| @@ -16,12 +16,12 @@ | |||
| from tokenizer.tiktoken import BaseTokenizer, IGNORE_INDEX | |||
| from torch.distributed.checkpoint.stateful import Stateful | |||
| from torch.utils.data import IterableDataset | |||
| from transform import CLIPTransform | |||
| from utils import load_image | |||
There was a problem hiding this comment.
NOTE: This is because of I ran pre-commit
| @@ -282,10 +282,12 @@ def __init__(self, moe_args: MoEArgs, dim: int, hidden_dim: int): | |||
| self.register_buffer( | |||
| "expert_bias", | |||
| torch.zeros(num_experts, dtype=torch.float32), | |||
| persistent=True, | |||
| ) | |||
There was a problem hiding this comment.
NOTE: Explicitly add whether the registered buffer is persistent. When false, we are not expected to load from DCP checkpoint.
tianyu-l
reviewed
Aug 12, 2025
There was a problem hiding this comment.
If I understand correctly, the conversion
- can be used to offline convert HF checkpoint from fp8 to fp32 using CPU plain tensor.
- can't be used to convert HF checkpoint on the fly using GPU DTensor, because of sharding and quantized blocks may not be aligned well.
- can't be used for weight sync to generate a state dict of bf16 because fake quantization to fp8 is applied.
I think it's OK to land this PR to unblock 1, but better to explain things clearly somewhere.
I also had some inline comments.
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Outdated suggestions cannot be applied.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.
Support loading a DeepSeek HF weights to Deepseek-V3 model:
Numerical verification: (using offline conversion script)