Skip to content

[Bugfix] Fix broken MTP weight loading for FP8 KV Scales #27227

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
Oct 21, 2025
Merged

[Bugfix] Fix broken MTP weight loading for FP8 KV Scales #27227

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
9 changes: 8 additions & 1 deletion vllm/model_executor/models/deepseek_mtp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,10 @@
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors

Expand Down Expand Up @@ -278,6 +281,10 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
if name.endswith(".bias") and name not in params_dict:
continue

name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
continue

# According to DeepSeek-V3 Technical Report, MTP modules
# shares embedding layer. We only load the first weights.
if (
Expand Down
Loading