[1/N][Feat] Xlite Qwen3 MoE Support

[changdawei1]

What this PR does / why we need it?

This patch adds support for the Qwen3-MoE model in Xlite. For more details about Xlite, please refer to the following link:https://atomgit.com/openeuler/GVirt/blob/master/xlite/README.md.

Qwen3-MoE TODO List:

• Qwen3-235B-A22B support
• Qwen3-MoE weights NZ support
• Qwen3-MoE data parallel support

Qwen3-30B-A3B-Instruct-2507 910B3(A2) Online Inference Performance Comparison

• aclgraph: main(69b170b)
• xlite-full: main + xlite-full
• xlite-decode-only: main + xlite-decode-only
• diff1: Performance comparison between xlite-full and aclgraph
• diff2: Performance comparison between xlite-decode-only and aclgraph

maxconcurrency	item	TTFT(ms)		TPOT(ms)		QPS (req/s)	OutputSpeed (token/s)
		Avg	P99	Avg	P99
1	baseline-aclgraph	205.07	287.29	12.34	12.65	0.14	78.81
1	xlite-full	66.40	113.69	11.71	12.40	0.15	84.73
1	xlite-decode-only	221.15	316.40	12.16	12.91	0.14	79.70
1	diff1	-67.62%	-60.43%	-5.11%	-1.98%	7.14%	7.51%
1	diff2	7.84%	10.13%	-1.46%	2.06%	0.00%	1.13%
16	baseline-aclgraph	1892.16	13916.86	22.78	39.28	1.15	589.89
16	xlite-full	1355.40	8907.45	15.96	25.15	1.65	850.21
16	xlite-decode-only	1519.42	8711.64	19.23	29.73	1.38	711.60
16	diff1	-28.37%	-36.00%	-29.94%	-35.97%	43.48%	44.13%
16	diff2	-19.70%	-37.40%	-15.58%	-24.31%	20.00%	20.63%
32	baseline-aclgraph	673.80	3914.90	32.20	37.95	1.80	928.54
32	xlite-full	481.65	2710.50	19.95	25.35	2.91	1506.67
32	xlite-decode-only	372.22	1095.25	25.19	28.47	2.33	1202.82
32	diff1	-28.52%	-30.76%	-38.04%	-33.20%	61.67%	62.26%
32	diff2	-44.76%	-72.02%	-21.77%	-24.98%	29.44%	29.54%
48	baseline-aclgraph	583.18	3277.65	41.02	46.05	2.17	1115.08
48	xlite-full	973.42	8237.33	23.29	30.50	3.71	1908.09
48	xlite-decode-only	480.79	2026.98	31.48	35.41	2.83	1453.75
48	diff1	66.92%	151.32%	-43.22%	-33.77%	70.97%	71.12%
48	diff2	-17.56%	-38.16%	-23.26%	-23.11%	30.41%	30.37%
64	baseline-aclgraph	742.74	5953.39	47.79	53.15	2.48	1272.37
64	xlite-full	545.22	3941.34	25.09	30.41	4.64	2376.44
64	xlite-decode-only	752.40	4534.29	38.67	43.28	3.06	1567.94
64	diff1	-26.59%	-33.80%	-47.50%	-42.78%	87.10%	86.77%
64	diff2	1.30%	-23.84%	-19.08%	-18.57%	23.39%	23.23%
100	baseline-aclgraph	565.52	1716.81	60.89	68.69	3.08	1580.64
100	xlite-full	398.14	2328.88	30.70	32.45	6.01	3086.42
100	xlite-decode-only	712.53	4875.94	52.71	60.78	3.53	1813.58
100	diff1	-29.60%	35.65%	-49.58%	-52.76%	95.13%	95.26%
100	diff2	26.00%	184.01%	-13.43%	-11.52%	14.61%	14.74%
150	baseline-aclgraph	842.42	5175.01	73.60	88.18	3.80	1952.26
150	xlite-full	568.52	4204.33	37.90	40.01	7.27	3734.72
150	xlite-decode-only	654.43	2504.06	67.40	77.00	4.18	2145.11
150	diff1	-32.51%	-18.76%	-48.51%	-54.63%	91.32%	91.30%
150	diff2	-22.32%	-51.61%	-8.42%	-12.68%	10.00%	9.88%
200	baseline-aclgraph	750.63	3049.91	88.26	101.95	4.28	2189.72
200	xlite-full	558.48	3791.98	45.54	49.04	8.17	4175.52
200	xlite-decode-only	807.09	4254.95	85.18	101.79	4.44	2271.52
200	diff1	-25.60%	24.33%	-48.40%	-51.90%	90.89%	90.69%
200	diff2	7.52%	39.51%	-3.49%	-0.16%	3.74%	3.74%

online server config:

port=$1
log=$2
export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1
export HCCL_BUFFSIZE=512
export HCCL_OP_EXPANSION_MODE="AIV"
export OMP_PROC_BIND=false
export VLLM_ASCEND_ENABLE_NZ=0
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000
ip=127.0.0.1
python -m vllm.entrypoints.openai.api_server \
	--model /mnt/nvme0n1/models/Qwen3-30B-A3B-Instruct-2507  \
	--tensor-parallel-size 8 \
	--gpu-memory-utilization 0.9 \
	--max-num-batched-tokens 32768 \
	--max-num-seqs=200 \
	--block-size 128 \
	--max-model-len 6656 \
	--trust-remote-code \
	--disable-log-requests \
	--served-model-name qwen \
	--no-enable-prefix-caching \
	--additional-config '{"xlite_graph_config": {"enabled": true, "full_mode": true}, "enable_cpu_binding": true}' \
	--compilation-config '{"cudagraph_capture_sizes":[1, 16, 32, 48, 64, 100, 150, 200], "cudagraph_mode": "FULL_DECODE_ONLY"}' \
	--async-scheduling \
	--host ${ip} \
	--port ${port} > ${log} 2>&1 &

test_config:

vllm bench serve \
    --max-concurrency ${maxconcurrency} \
    --num-prompts ${num_prompts} \
    --host ${HOST} \
    --port ${PORT} \
    --model ${MODEL_NAME} \
    --dataset-name random \
    --backend openai-chat \
    --random-input-len 512 \
    --random-output-len 512  \
    --random-range-ratio 0.2 \
    --temperature 0.6 \
    --metric-percentiles "50,90,99" \
    --tokenizer ${TOKENIZER_PATH} \
    --endpoint /v1/chat/completions \
    --ignore-eos

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.13.0
• vLLM main: vllm-project/vllm@2c24bc6

[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 adds support for Qwen3 MoE models in Xlite graph mode. The changes include updating documentation, dependencies, and tests. The core logic is in vllm_ascend/xlite/xlite.py, where a new QwenMoeXliteModel class is introduced to handle MoE-specific configurations and model building. The existing LlamaXliteModel has been refactored to facilitate this extension. My review found a critical issue in the new MoE model configuration logic that could lead to a crash when expert parallelism is disabled. A fix has been suggested.

[gemini-code-assist]

The method _build_model_config in QwenMoeXliteModel unconditionally calls get_ep_group() on line 208. This will cause a crash with an AssertionError if expert parallelism is not enabled, as the expert parallel group (_EP_GROUP) will not be initialized. The call to get_ep_group() should be conditional on vllm_config.parallel_config.enable_expert_parallel being true. Additionally, the logic for setting moe_tp_size in the else branch is incorrect as it also relies on ep_group which would not be available.

    def _build_model_config(self, vllm_config: VllmConfig) -> ModelConfig:
        config = super()._build_model_config(vllm_config)
        hf_config = vllm_config.model_config.hf_text_config
        config.n_layers = hf_config.max_window_layers
        config.n_dense_layers = 0
        config.n_routed_experts = hf_config.num_experts
        config.n_shared_experts = 0
        config.n_act_experts = hf_config.num_experts_per_tok
        config.def_dp_size = vllm_config.parallel_config.data_parallel_size
        if vllm_config.parallel_config.enable_expert_parallel:
            ep_group = get_ep_group()
            config.moe_ep_size = ep_group.world_size
            config.moe_tp_size = 1
        else:
            config.moe_ep_size = 1
            config.moe_tp_size = 1
        config.experts_weight_transpose = True
        config.moe_intermediate_size = hf_config.moe_intermediate_size
        config.norm_topk_prob = hf_config.norm_topk_prob
        config.scoring_func = ScoringFuncSoftmax
        return config

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.