Skip to content

Revert "Embedding parallel by attn_tp (#7623)" #7880

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jul 9, 2025
Merged

Revert "Embedding parallel by attn_tp (#7623)" #7880

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion python/sglang/srt/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1938,7 +1938,7 @@ def __init__(
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
use_attn_tp_group=True,
enable_tp=not global_server_args_dict["enable_dp_attention"],
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jul 9, 2025

Choose a reason for hiding this comment

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

medium

For better maintainability, consider adding a comment explaining why tensor parallelism for embeddings is disabled when data-parallel attention is enabled. This clarifies the rationale behind the logic.

# When data-parallel attention is enabled, the input batch is split across DP ranks.
# To avoid complex communication for the embedding lookup, we replicate the embedding layer by disabling its tensor parallelism.
self.embed_tokens = VocabParallelEmbedding(
    config.vocab_size,
    config.hidden_size,
    enable_tp=not global_server_args_dict["enable_dp_attention"],
)

)
self.alt_stream = torch.cuda.Stream() if _is_cuda else None
self.layers = nn.ModuleList(
Expand Down
Loading