Skip to content

[Bugfix] Pad hidden_states to avoid cross-ring AllGatherV #963

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[Bugfix] Pad hidden_states to avoid cross-ring AllGatherV #963

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bb6ea4e
[Bugfix] Pad hidden_states to avoid cross-ring AllGatherV
ApsarasX May 26, 2025
cc7ec02
[Chore] Rename padded_num_tokens to num_padding_tokens for consistency
ApsarasX May 27, 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
23 changes: 10 additions & 13 deletions vllm_ascend/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,21 +222,18 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
else:
is_prefill = attn_metadata.num_prefills > 0
enable_force_load_balance = False
num_tokens, hidden_dim = hidden_states.shape
num_tokens, hidden_size = hidden_states.shape

if self.n_shared_experts is not None:
shared_output = self.shared_experts(hidden_states)

if self.tp_size > 1:
# pass
num_tokens, hidden_size = hidden_states.shape
if num_tokens < self.tp_size:
target_size = self.tp_size
new_hidden_states = torch.empty([target_size, hidden_size],
dtype=hidden_states.dtype,
device=hidden_states.device)
new_hidden_states[:num_tokens] = hidden_states
hidden_states = new_hidden_states
num_padding_tokens = (self.tp_size -
num_tokens % self.tp_size) % self.tp_size
# Pad hidden_states to make it divisible by tp_size to avoid cross-ring AllGatherV on 910B2C
if num_padding_tokens > 0:
hidden_states = nn.functional.pad(
hidden_states, (0, 0, 0, num_padding_tokens))
chunk_hidden_states = torch.tensor_split(hidden_states,
self.tp_size,
dim=0)
Expand All @@ -259,15 +256,15 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
dist.all_gather(list(chunk_hidden_states), router_hidden_states,
self.tp_group)
final_hidden_states = torch.cat(chunk_hidden_states, dim=0)
if num_tokens < self.tp_size:
final_hidden_states = final_hidden_states[:num_tokens]
if num_padding_tokens > 0:
final_hidden_states = final_hidden_states[:-num_padding_tokens]
else:
final_hidden_states = router_hidden_states

if shared_output is not None:
final_hidden_states = final_hidden_states + shared_output

return final_hidden_states.view(num_tokens, hidden_dim)
return final_hidden_states.view(num_tokens, hidden_size)


class CustomDeepseekV2MLAAttention(DeepseekV2MLAAttention):
Expand Down