Skip to content

[Optimization] EP empty_input_forward Remove Communication #5254

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
ckl117 merged 6 commits into from
Dec 1, 2025
Merged

[Optimization] EP empty_input_forward Remove Communication #5254

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
8ac7918
check fake input shape
ckl117 Nov 26, 2025
f1d7bc5
Merge branch 'develop' of https://github.com/PaddlePaddle/FastDeploy …
ckl117 Nov 28, 2025
5d2ba1c
support glm 4.6
ckl117 Nov 28, 2025
4a3a8b8
Merge branch 'develop' of https://github.com/PaddlePaddle/FastDeploy …
ckl117 Nov 30, 2025
2054605
check
ckl117 Dec 1, 2025
3169f0c
Merge branch 'develop' of https://github.com/PaddlePaddle/FastDeploy …
ckl117 Dec 1, 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
5 changes: 5 additions & 0 deletions custom_ops/gpu_ops/moe/ep_moe_expert_dispatch.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,6 +58,11 @@
__VA_ARGS__ \
break; \
} \
case 20: { \
constexpr size_t NUM_EXPERTS_PER_RANK = 20; \
__VA_ARGS__ \
break; \
} \
case 32: { \
constexpr size_t NUM_EXPERTS_PER_RANK = 32; \
__VA_ARGS__ \
Expand Down
4 changes: 3 additions & 1 deletion fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,8 +146,10 @@ def apply_ep_prefill(
recv_topk_weights,
recv_num_tokens_per_expert_list,
handle,
_,
event,
) = self.ep_prefill_runner.dispatch(x, topk_idx, topk_weights)
if self.ep_prefill_runner.ep_engine.async_finish:
event.current_stream_wait()
token_all_num = sum(recv_num_tokens_per_expert_list)

# 3. Compute ffn
Expand Down
4 changes: 2 additions & 2 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 @@ -498,8 +498,8 @@ def empty_input_forward(self):
"""
empty_input_forward
"""
fake_hidden_states = paddle.ones(
shape=[1, self.fd_config.model_config.hidden_size],
fake_hidden_states = paddle.empty(
shape=[0, self.fd_config.model_config.hidden_size],
dtype=paddle.get_default_dtype(),
)
for i in range(
Expand Down
2 changes: 1 addition & 1 deletion fastdeploy/model_executor/models/qwen3moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -421,7 +421,7 @@ def empty_input_forward(self):
empty_input_forward
"""
fake_hidden_states = paddle.empty(
shape=[1, self.fd_config.model_config.hidden_size],
shape=[0, self.fd_config.model_config.hidden_size],
Copy link
Collaborator

@K11OntheBoat K11OntheBoat Dec 1, 2025

Choose a reason for hiding this comment

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

LGTM

dtype=paddle.get_default_dtype(),
)
for i in range(
Expand Down
Loading