Skip to content

[Cherry-Pick][Optimization] merge_allreduce (#7039) #7475

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

Draft
Copilot wants to merge 1 commit into
base: release/2.5
Choose a base branch
from
Draft

[Cherry-Pick][Optimization] merge_allreduce (#7039) #7475

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
17 changes: 14 additions & 3 deletions fastdeploy/model_executor/models/glm4_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@

import fastdeploy
from fastdeploy.config import FDConfig
from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
from fastdeploy.model_executor.forward_meta import ForwardMeta
from fastdeploy.model_executor.graph_optimization.decorator import (
support_graph_optimization,
Expand Down Expand Up @@ -158,8 +159,16 @@ def __init__(
default_initializer=paddle.nn.initializer.Constant(0),
)

# In pure-TP mode (tp>1, ep=1) both branches return partial sums, so we
# defer the all-reduce to after combining them — saving one collective.
# In all other modes (EP, EP+attn-TP, no parallelism) each branch handles
# its own reduction internally (reduce_results default=True), so we must
# NOT add an extra all-reduce here.
self.merge_ffn_tp = self.use_tp and not self.use_ep

self.experts = FusedMoE(
fd_config,
reduce_results=not self.merge_ffn_tp,
renormalize=self.norm_topk_prob,
moe_intermediate_size=fd_config.model_config.moe_intermediate_size,
num_experts=fd_config.model_config.n_routed_experts,
Expand All @@ -181,14 +190,16 @@ def __init__(
intermediate_size=shared_experts_intermediate_size,
layer_id=layer_id,
prefix=f"{prefix}.shared_experts",
reduce_results=not self.merge_ffn_tp,
)

def forward(self, x, forward_meta: ForwardMeta = None):
out = self.experts(x, self.gate, forward_meta)
if self.n_shared_experts > 0:
shared_experts_out = self.shared_experts(x)
out = out + shared_experts_out

out = out + self.shared_experts(x)
if self.merge_ffn_tp:
# Both branches produced partial sums; combine first, then single all-reduce.
out = tensor_model_parallel_all_reduce(out, self.tp_group)
return out


Expand Down