Skip to content

[RL] Allow passing tensors of different dtypes for FlattenedTensorBucket #13413

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
fzyzcjy merged 5 commits into from
Nov 19, 2025
Merged

[RL] Allow passing tensors of different dtypes for FlattenedTensorBucket #13413

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e18467a
[RL] Allow passing tensors of different dtypes for FlattenedTensorBucket
zhuzilin Nov 17, 2025
8f7b1a2
add supports_multi_dtypes key for backward compatibility
zhuzilin Nov 17, 2025
474d95d
Apply suggestion from @fzyzcjy
fzyzcjy Nov 17, 2025
f963760
Merge branch 'main' into feature/flatten_tensor_bucket
fzyzcjy Nov 17, 2025
ea36973
Merge branch 'main' into feature/flatten_tensor_bucket
fzyzcjy Nov 18, 2025
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
15 changes: 8 additions & 7 deletions python/sglang/srt/weight_sync/tensor_bucket.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,9 @@ class FlattenedTensorBucket:
while preserving all metadata needed for reconstruction.
"""

# This field is solely for users of to check whether the class supports this feature
supports_multi_dtypes = True

def __init__(
self,
named_tensors: List[Tuple[str, torch.Tensor]] = None,
Expand All @@ -48,7 +51,7 @@ def __init__(
flattened_tensors: List[torch.Tensor] = [None] * len(named_tensors)

for i, (name, tensor) in enumerate(named_tensors):
flattened = tensor.flatten()
flattened = tensor.flatten().view(torch.uint8)
flattened_tensors[i] = flattened

# Store metadata
Expand Down Expand Up @@ -93,14 +96,12 @@ def reconstruct_tensors(self) -> List[Tuple[str, torch.Tensor]]:
reconstructed = [None] * len(self.metadata)

for i, meta in enumerate(self.metadata):
tensor = self.flattened_tensor[meta.start_idx : meta.end_idx].reshape(
meta.shape
tensor = (
self.flattened_tensor[meta.start_idx : meta.end_idx]
.view(meta.dtype)
.reshape(meta.shape)
)

# batch dtype conversion (if needed)
if tensor.dtype != meta.dtype:
tensor = tensor.to(meta.dtype)

reconstructed[i] = (meta.name, tensor)

return reconstructed
Loading