[BugFix]fixed all_reduce_merge_allgather_ep bug

[ttanzhiqiang]

There are four situations in the previous logic

1. If Prefill/decode uses AllGather or NaiveMulticast at the same time, this logic is normal, and this solution is used for optimization
2. If Prefill/decode uses All2All/MC2 at the same time, this logic is also normal, and this solution is not used for optimization
3. Prefill uses AllGatherEP (using VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP switch), and Decode uses MC2. (bug)
4. In the PD separation scenario, the strategies used by P and D are separate, so there is no impact.

After modification

1. If Prefill/decode use AllGather or NaiveMulticast solution at the same time, this logic is normal, and this solution is used for optimization
2. If Prefill/decode use All2All/MC2 solution at the same time, this logic is also normal, and this solution is not used for optimization
3. Prefill uses AllGatherEP solution (use VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP switch), and Decode uses MC2 solution. (Prefill can be merged/Prefill and Decode strategies are different and cannot be merged)
4. In the PD separation scenario, the strategies used by P and D are separate, so there is no impact.

How was this patch tested?

Test method for case 1

export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export ASCEND_LAUNCH_BLOCKING=0
export VLLM_VERSION=0.9.1
nohup python -m vllm.entrypoints.openai.api_server --model=/mnt/deepseek/DeepSeek-R1-W8A8-VLLM
--served-model-name auto
--quantization ascend
--trust-remote-code
--distributed-executor-backend=mp
--port 8006
-tp=4
-dp=4
--max-num-seqs 24
--max-model-len 32768
--max-num-batched-tokens 32768
--block-size 128
--no-enable-prefix-caching
--additional-config '{"torchair_graph_config":{"enabled":true,"use_cached_graph":true,"graph_batch_sizes":[24]},"ascend_scheduler_config":{"enabled":true},"expert_tensor_parallel_size":16}'
--gpu-memory-utilization 0.96 &> run.log &
disown

Test method for case 2

export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export ASCEND_LAUNCH_BLOCKING=0
export VLLM_VERSION=0.9.1
nohup python -m vllm.entrypoints.openai.api_server --model=/mnt/deepseek/DeepSeek-R1-W8A8-VLLM
--served-model-name auto
--quantization ascend
--trust-remote-code
--distributed-executor-backend=mp
--port 8006
-tp=4
-dp=4
--max-num-seqs 24
--max-model-len 32768
--max-num-batched-tokens 32768
--block-size 128
--no-enable-prefix-caching
--enable_expert_parallel
--additional-config '{"torchair_graph_config":{"enabled":true,"use_cached_graph":true,"graph_batch_sizes":[24]},"ascend_scheduler_config":{"enabled":true},"expert_tensor_parallel_size":16}'
--gpu-memory-utilization 0.96 &> run.log &
disown

Test method for case 3

export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export ASCEND_LAUNCH_BLOCKING=0
export VLLM_VERSION=0.9.1
export VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP=1
nohup python -m vllm.entrypoints.openai.api_server --model=/mnt/deepseek/DeepSeek-R1-W8A8-VLLM
--served-model-name auto
--quantization ascend
--trust-remote-code
--distributed-executor-backend=mp
--port 8006
-tp=4
-dp=4
--max-num-seqs 24
--max-model-len 32768
--max-num-batched-tokens 32768
--block-size 128
--no-enable-prefix-caching
--enable_expert_parallel
--additional-config '{"torchair_graph_config":{"enabled":true,"use_cached_graph":true,"graph_batch_sizes":[24]},"ascend_scheduler_config":{"enabled":true},"expert_tensor_parallel_size":16}'
--gpu-memory-utilization 0.96 &> run.log &
disown

• vLLM version: v0.9.2
• vLLM main: vllm-project/vllm@e7e3e6d

[codecov]

Codecov Report

Attention: Patch coverage is 16.66667% with 5 lines in your changes missing coverage. Please review.

Project coverage is 53.49%. Comparing base (f96100f) to head (3e02626).

Files with missing lines	Patch %	Lines
vllm_ascend/models/deepseek_v2.py	25.00%	3 Missing ⚠️
vllm_ascend/ops/fused_moe.py	0.00%	2 Missing ⚠️

❌ Your patch check has failed because the patch coverage (16.66%) is below the target coverage (100.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1818      +/-   ##
==========================================
- Coverage   53.51%   53.49%   -0.02%     
==========================================
  Files          77       77              
  Lines        9435     9440       +5     
==========================================
+ Hits         5049     5050       +1     
- Misses       4386     4390       +4     

Flag	Coverage Δ
unittests	53.49% <16.66%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull Request Overview

This PR refines the all_reduce_merge behavior to correctly handle the case where Prefill uses the AllGatherEP solution and Decode uses the MC2 solution, ensuring merge happens only in Prefill.

• Expanded comments in get_all_reduce_merge_state to detail four merge scenarios
• Introduced a conditional all-reduce for shared experts in fused_moe.py when Prefill uses AllGatherEP and Decode uses MC2
• Added a stub branch in deepseek_v2.py for the new AllGatherEP-then-MC2 path

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
vllm_ascend/utils.py	Clarified four scenarios for all_reduce_merge in doc comments
vllm_ascend/ops/fused_moe.py	Added conditional merge for shared_hidden_states under EP+MC2 path
vllm_ascend/models/deepseek_v2.py	Imported envs_ascend and added placeholder for EP+MC2 logic

Comments suppressed due to low confidence (3)

vllm_ascend/utils.py:461

• [nitpick] The inline comment list has inconsistent punctuation and phrasing across items. Consider ending each item with a period and using parallel sentence structure for better readability.

# 1. If Prefill/decode use AllGather or NaiveMulticast solution at the same time, this logic is normal, and this solution is used for optimization

vllm_ascend/ops/fused_moe.py:1421

• The code references envs_ascend but there is no import of vllm_ascend.envs as envs_ascend at the top of this file, which will raise a NameError at runtime. Please add the appropriate import.

        if tp_size > 1 and envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and self.all_reduce_merge and fused_moe_state in [

vllm_ascend/models/deepseek_v2.py:413

• The ... placeholder in this branch is a syntax stub and will cause a runtime error. Implement the intended merge logic or add a clear raise NotImplementedError and a TODO.

                ...

[Copilot]

[nitpick] For a single-state check, using fused_moe_state == FusedMoEState.MC2 is clearer and more efficient than membership in a one-element list.

Suggested change

	if tp_size > 1 and envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and self.all_reduce_merge and fused_moe_state in [
	FusedMoEState.MC2
	]:
	if tp_size > 1 and envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and self.all_reduce_merge and fused_moe_state == FusedMoEState.MC2: