Skip to content

[BugFix] Partial revert of #29558 (DeepEP HT + PIECEWISE CG support) #30910

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
khluu merged 6 commits into from
Dec 18, 2025
Merged

[BugFix] Partial revert of #29558 (DeepEP HT + PIECEWISE CG support) #30910

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
18986f4
partial revert
LucasWilkinson Dec 17, 2025
ef34bf9
wip
LucasWilkinson Dec 17, 2025
417e0dc
Update vllm/config/compilation.py
LucasWilkinson Dec 17, 2025
9388297
remove added
LucasWilkinson Dec 17, 2025
ee7bd13
cleanup
LucasWilkinson Dec 17, 2025
4f1fca1
Merge branch 'main' into lwilkinson/partial-revert
LucasWilkinson Dec 18, 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
38 changes: 0 additions & 38 deletions tests/compile/test_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -233,24 +233,6 @@ def test_splitting_ops_dynamic():
assert config.compilation_config.cudagraph_mode == CUDAGraphMode.PIECEWISE


def test_moe_splitting_ops_deepep_ht_piecewise():
# Non-inductor, non-attn-fusion case: DeepEP HT with dp>1
# should add MoE ops to splitting_ops on top of attention ops.
config = VllmConfig(
parallel_config=ParallelConfig(
all2all_backend="deepep_high_throughput",
data_parallel_size=8,
),
compilation_config=CompilationConfig(
mode=CompilationMode.VLLM_COMPILE,
),
)
splitting_ops = config.compilation_config.splitting_ops
assert splitting_ops is not None
assert "vllm::moe_forward" in splitting_ops
assert "vllm::moe_forward_shared" in splitting_ops


def test_moe_splitting_ops_deepep_ht_inductor_partition():
# Inductor partition case: user-provided splitting_ops should be
# preserved and MoE ops should be appended for DeepEP HT with dp>1.
Expand All @@ -277,26 +259,6 @@ def test_moe_splitting_ops_deepep_ht_inductor_partition():
]


def test_moe_splitting_ops_deepep_ht_attn_fusion_no_inductor():
# Pure attn-fusion case without inductor partition: even with
# DeepEP HT and dp>1, we should not re-enable piecewise compilation
# or add MoE ops into splitting_ops.
config = VllmConfig(
parallel_config=ParallelConfig(
all2all_backend="deepep_high_throughput",
data_parallel_size=8,
),
compilation_config=CompilationConfig(
mode=CompilationMode.VLLM_COMPILE,
pass_config={"fuse_attn_quant": True, "eliminate_noops": True},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
),
)
assert config.compilation_config.splitting_ops == []
assert config.compilation_config.cudagraph_mode == CUDAGraphMode.FULL


def test_should_split():
import torch

Expand Down
48 changes: 15 additions & 33 deletions vllm/config/compilation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -930,7 +930,7 @@
# for details. Make a copy to avoid mutating the class-level
# list via reference.
self.splitting_ops = list(self._attention_ops)
added_default_splitting_ops = True

Check failure on line 933 in vllm/config/compilation.py

View workflow job for this annotation

GitHub Actions / pre-commit

Ruff (F841) 

vllm/config/compilation.py:933:17: F841 Local variable `added_default_splitting_ops` is assigned to but never used
elif len(self.splitting_ops) == 0:
if (
self.cudagraph_mode == CUDAGraphMode.PIECEWISE
Expand Down Expand Up @@ -958,44 +958,26 @@
self.cudagraph_mode = CUDAGraphMode.FULL
self.splitting_ops = []

# split MoE ops for cudagraph
moe_ops = [
"vllm::moe_forward",
"vllm::moe_forward_shared",
]
# Disable CUDA graphs for DeepEP high-throughput with DP > 1
backend = all2all_backend or envs.VLLM_ALL2ALL_BACKEND
dp_size = data_parallel_size if data_parallel_size is not None else 1
need_moe_splitting = (
if (
backend == "deepep_high_throughput"
and dp_size > 1
# pure attn-fusion without inductor partition deliberately disables
# piecewise graphs and MoE splitting.
and not (
self.pass_config.fuse_attn_quant
and not self.use_inductor_graph_partition
and self.cudagraph_mode != CUDAGraphMode.NONE
):
# TODO: Piecewise Cuda graph might be enabled
# if torch compile cache key issue fixed
# See https://github.com/vllm-project/vllm/pull/25093
logger.info(
"WideEP: Disabling CUDA Graphs since DeepEP high-throughput "
Comment thread
LucasWilkinson marked this conversation as resolved.
Outdated
"kernels are optimized for prefill and are incompatible with "
"CUDA Graphs. "
"In order to use CUDA Graphs for decode-optimized workloads, "
"use --all2all-backend with another option, such as "
"deepep_low_latency, pplx, or allgather_reducescatter."
)
)

if need_moe_splitting and self.cudagraph_mode != CUDAGraphMode.NONE:
# if we just initialized default splitting_ops for this config,
# automatically append the MoE ops
if added_default_splitting_ops:
for op in moe_ops:
if op not in self.splitting_ops:
self.splitting_ops.append(op)

# make sure MoE ops are split out
if not any(op in self.splitting_ops for op in moe_ops):
self.cudagraph_mode = CUDAGraphMode.NONE
logger.warning_once(
"DeepEP high throughput backend with data_parallel_size > 1 "
"requires splitting MoE ops from cudagraphs. Please ensure "
"'vllm::moe_forward' or 'vllm::moe_forward_shared' are "
"present in CompilationConfig.splitting_ops."
)
elif self.cudagraph_mode.has_full_cudagraphs():
# fall back to piecewise when MoE splitting is required.
self.cudagraph_mode = CUDAGraphMode.PIECEWISE
self.cudagraph_mode = CUDAGraphMode.NONE

def set_splitting_ops_for_attn_fusion(self):
assert self.pass_config.fuse_attn_quant
Expand Down
Loading