Skip to content

[DeepEP] Eliminate unnecessary DP cudagraph padding #5557

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 3 commits into from
+11 −2
Closed

[DeepEP] Eliminate unnecessary DP cudagraph padding #5557

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a15d7cd
[DeepEP] Eliminate unnecessary DP cudagraph padding
yuleil Apr 19, 2025
349df6d
Padding is necessary when moe_dense_tp_size!=1
yuleil Apr 23, 2025
6280e6b
typo
yuleil Apr 23, 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
13 changes: 11 additions & 2 deletions python/sglang/srt/model_executor/cuda_graph_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -188,6 +188,11 @@ def __init__(self, model_runner: ModelRunner):
self.speculative_algorithm = model_runner.server_args.speculative_algorithm
self.tp_size = model_runner.server_args.tp_size
self.dp_size = model_runner.server_args.dp_size
self.enable_deepep_moe = model_runner.server_args.enable_deepep_moe
self.moe_dense_fully_dp = (
model_runner.server_args.enable_deepep_moe
and model_runner.server_args.moe_dense_tp_size == 1
)

# Batch sizes to capture
self.capture_bs, self.compile_bs = get_batch_sizes_to_capture(model_runner)
Expand Down Expand Up @@ -301,7 +306,9 @@ def model_capture_mode(self):

def can_run(self, forward_batch: ForwardBatch):
if self.enable_dp_attention or self.enable_sp_layernorm:
total_global_tokens = sum(forward_batch.global_num_tokens_cpu)
reducer = max if self.moe_dense_fully_dp else sum
# DeepEP MoE layers uses a fixed shape with masking instead of gather tokens from DP ranks.
total_global_tokens = reducer(forward_batch.global_num_tokens_cpu)

is_bs_supported = forward_batch.can_run_dp_cuda_graph and (
total_global_tokens in self.graphs
Expand Down Expand Up @@ -484,8 +491,10 @@ def replay_prepare(self, forward_batch: ForwardBatch):

# Pad
if self.enable_dp_attention or self.enable_sp_layernorm:
reducer = max if self.moe_dense_fully_dp else sum
# DeepEP MoE layers uses a fixed shape with masking instead of gather tokens from DP ranks.
index = bisect.bisect_left(
self.capture_bs, sum(forward_batch.global_num_tokens_cpu)
self.capture_bs, reducer(forward_batch.global_num_tokens_cpu)
)
else:
index = bisect.bisect_left(self.capture_bs, raw_bs)
Expand Down
Loading