Skip to content

[bugfix]fix multistream moe in torchair #3164

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 1 commit into
base: main
Choose a base branch
from
+4 −2
Open

[bugfix]fix multistream moe in torchair #3164

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
6 changes: 4 additions & 2 deletions vllm_ascend/torchair/models/torchair_deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,12 +160,13 @@ def forward(
output_parallel = self.quant_method.apply(self,
input_parallel,
bias=bias_)
forward_context = get_forward_context()
if self.reduce_results and self.tp_size > 1:
num_tokens = output_parallel.shape[0]
if is_force_scatter and num_tokens % self.tp_size:
output_parallel = nn.functional.pad(
output_parallel, (0, 0, 0, -num_tokens % self.tp_size))
if is_force_scatter or (not is_prefill
if is_force_scatter or (not forward_context.with_prefill
and output_parallel.shape[0] % self.tp_size
== 0):
Comment on lines +169 to 171
Copy link
Contributor

Choose a reason for hiding this comment

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

high

Using forward_context.with_prefill makes the is_prefill parameter of this forward method unused. This is misleading and can lead to future bugs if other developers rely on this parameter, expecting it to have an effect.

To improve maintainability and prevent potential issues, the is_prefill parameter should be removed from the function signature of TorchairDeepseekV2RowParallelLinearReplaceAllreduce.forward. This will require updating the call site in vllm_ascend/attention/mla_v1.py to no longer pass this argument.

output = tensor_model_parallel_reduce_scatter(output_parallel,
Expand Down Expand Up @@ -726,7 +727,8 @@ def forward(
replace_allreduce: bool = False,
) -> torch.Tensor:
# Self Attention
if attn_metadata is not None and attn_metadata.num_decodes > 0:
forward_context = get_forward_context()
if attn_metadata is not None and not forward_context.with_prefill:
mla_moe_communication = self.mla_moe_communication and replace_allreduce
else:
mla_moe_communication = False
Expand Down
Loading