[CustomOp] Register AscendSharedFusedMoE custom op

[shen-shanshan]

What this PR does / why we need it?

Register AscendSharedFusedMoE custom op.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

DeepSeek-V2-Lite is a MoE model with shared experts.

Test:

vllm serve /root/.cache/modelscope/hub/models/deepseek-ai/DeepSeek-V2-Lite \
--trust-remote-code \
--enforce-eager \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.95

curl -X POST http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "/root/.cache/modelscope/hub/models/deepseek-ai/DeepSeek-V2-Lite",
        "messages": [
            {"role": "user", "content": "介绍一下联通公司？"}
        ],
        "stream": false,
        "max_tokens": 100
    }'

Output:

中国联合网络通信集团有限公司（简称“中国联通”）于2009年1月6日在原中国网通和原中国联通的基础上合并组建而成，在国内31个省（自治区、直辖市）和境外多个国家和地区设有分支机构，是中国唯一一家在纽约、香港、上海三地同时上市的电信运营企业，连续多年入选“世界500强企业”。\n\n中国联通主要经营固定通信业务，移动通信业务，国内

• vLLM version: v0.10.2
• vLLM main: vllm-project/vllm@486c559

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request refactors the registration of the AscendSharedFusedMoE custom operator by replacing a model-specific monkey-patch with a centralized, global registration. This is a cleaner approach architecturally. However, I've raised one high-severity concern regarding the change in scope from a specific patch to a global one, which could introduce regressions in models that are not expecting this specific implementation of SharedFusedMoE. Apart from this risk, the implementation of the change is straightforward.

[yiz-liu]

I don't quite get it—why do we need forward_impl? Why don't we just call AscendFusedMoE.forward_impl() directly from AscendSharedFusedMoE.forward?

[shen-shanshan]

@yiz-liu

Calling Workflow:

1. AscendSharedFusedMoE: forward()
2. AscendFusedMoE: forward()
3. CustomOp: forward()
4. CustomOp: forward_oot()
5. FusedMoE: forward_native()
6. AscendSharedFusedMoE: forward_impl()
7. AscendFusedMoE: forward_impl()

In AscendFusedMoE: forward_impl() (step 7), we just return fused_output, but in FusedMoE forward_native() (step 5), it recieve 2 results (shared_output and fused_output), which is conflict with line 443 in AscendSharedFusedMoE.

Thus, we need to return a dummy shared_output in forward_impl() of AscendSharedFusedMoE.

[shen-shanshan]

See https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/layer.py#L1729.

[shen-shanshan]

Test GLM-4.5

Run:

export VLLM_USE_V1=1
export VLLM_WORKER_MULTIPROC_METHOD="spawn"
export VLLM_USE_MODELSCOPE=True
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=100
export HCCL_BUFFSIZE=1024
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256

vllm serve /root/.cache/modelscope/hub/models/ZhipuAI/GLM-4___5 \
--host 0.0.0.0 \
--port 8004 \
--tensor-parallel-size 16 \
--max-num-seqs 8 \
--max-model-len 16384 \
--max-num-batched-tokens 2048 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.95 \
--additional-config '{"ascend_scheduler_config":{"enabled":true, "enable_chunked_prefill": true}}' \
--enable-expert-parallel \
--enforce-eager

Curl:

curl -X POST http://172.22.0.155:8004/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "/root/.cache/modelscope/hub/models/ZhipuAI/GLM-4___5",
        "messages": [
            {"role": "user", "content": "介绍一下联通公司？"}
        ],
        "stream": false,
        "max_tokens": 100
    }'

Output:

\n<think>嗯，用户让我介绍一下联通公司。首先得确定用户需要哪些方面的信息。可能用户只是想了解基本概况，比如成立时间、主要业务，或者也可能对联通的市场地位、技术发展感兴趣。\n\n接下来，我需要回忆联通的基础信息。成立于1994年，原称中国联通，2009年和中国网通合并，改名为中国联通。这点要明确，避免混淆。然后总部在北京，央企属性很重要，说明其背景和稳定性。\n\n业务方面

[yiz-liu]

@shen-shanshan This PR may have broken #2946 , please contact @whx-sjtu and see if you can work out a solution?