Skip to content

[V1][PP] Fix intermediate tensor values #13417

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
Feb 17, 2025
Merged

[V1][PP] Fix intermediate tensor values #13417

Show file tree
Hide file tree
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
3 changes: 3 additions & 0 deletions vllm/sequence.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1137,6 +1137,9 @@ def __getitem__(self, key: Union[str, slice]):
def __setitem__(self, key: str, value: torch.Tensor):
self.tensors[key] = value

def items(self):
return self.tensors.items()

def __len__(self):
return len(self.tensors)

Expand Down
10 changes: 8 additions & 2 deletions vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,7 +151,8 @@ def __init__(
self.positions = torch.zeros(self.max_num_tokens,
dtype=torch.int64,
device=self.device)
# self.intermediate_tensors # Set after load_model
# None in the first PP rank. The rest are set after load_model.
self.intermediate_tensors: Optional[IntermediateTensors] = None

# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
Expand Down Expand Up @@ -925,6 +926,11 @@ def execute_model(
if get_pp_group().is_first_rank:
intermediate_tensors = None
else:
assert intermediate_tensors is not None
assert self.intermediate_tensors is not None
for k, v in intermediate_tensors.items():
self.intermediate_tensors[k][:num_input_tokens].copy_(
v[:num_input_tokens], non_blocking=True)
intermediate_tensors = IntermediateTensors({
k: v[:num_input_tokens]
for k, v in self.intermediate_tensors.items()
Expand Down Expand Up @@ -1135,7 +1141,7 @@ def _dummy_run(
if get_pp_group().is_first_rank:
intermediate_tensors = None
else:
if not hasattr(self, "intermediate_tensors"):
if self.intermediate_tensors is None:
self.intermediate_tensors = (
self.model.make_empty_intermediate_tensors(
batch_size=self.max_num_tokens,
Expand Down