Skip to content

[Core/DBO][3/N] Dual-Batch Overlap with speculative decode #24608

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

Closed
wants to merge 1 commit into from
Closed

[Core/DBO][3/N] Dual-Batch Overlap with speculative decode #24608

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
18 changes: 16 additions & 2 deletions vllm/model_executor/layers/fused_moe/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1801,6 +1801,15 @@ def process_chunk(chunk_start, chunk_end, skip_result_store=False):
staged_hidden_states.copy_(hidden_states, non_blocking=True)
staged_router_logits.copy_(router_logits, non_blocking=True)

# If there are shared experts but we are not using a modular kernel,
# the shared experts must be called here
if (not isinstance(self.quant_method.fused_experts,
FusedMoEModularKernel)
and self.shared_experts is not None):
shared_output = self.shared_experts(staged_hidden_states)
else:
shared_output = None

# Matrix multiply.
final_hidden_states = self.quant_method.apply(
layer=self,
Expand All @@ -1824,8 +1833,13 @@ def process_chunk(chunk_start, chunk_end, skip_result_store=False):
logical_replica_count=self.logical_replica_count,
)

assert self.shared_experts is None or isinstance(
final_hidden_states, tuple)
if shared_output is not None:
assert not isinstance(final_hidden_states, tuple)
assert self.shared_experts is not None
final_hidden_states = (
shared_output,
final_hidden_states,
)

if not skip_result_store:
if self.shared_experts is None:
Expand Down