Skip to content

DeepSeek V4 RL support#28

Open
yueming-yuan wants to merge 2 commits into
radixark:miles-mainfrom
yueming-yuan:deepseek-v4
Open

DeepSeek V4 RL support#28
yueming-yuan wants to merge 2 commits into
radixark:miles-mainfrom
yueming-yuan:deepseek-v4

Conversation

@yueming-yuan
Copy link
Copy Markdown

No description provided.

Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>

# fp32 params.
elif model_param.type() == 'torch.cuda.FloatTensor':
# Keep shard tensors as leaf tensors for torch Optimizer.
Copy link
Copy Markdown

@guapisolo guapisolo Apr 24, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a fp32 bug in megatron I fixed before, so it shouldn’t negatively impact accuracy. No need to modify this.

@guapisolo
Copy link
Copy Markdown

guapisolo commented Apr 24, 2026

If fp32 precision code path need context, ping me anytime.

@yiakwy-xpu-ml-framework-team
Copy link
Copy Markdown

yiakwy-xpu-ml-framework-team commented Apr 29, 2026

@yueming-yuan have you performed a megatron fwd with a prompt ?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants