[Bug]: 使用vllm-ascend的v0.11.0rc1镜像跑不通v0.11.0rc1分支的一个脚本，但是0.9.1镜像能跑通，脚本路径：https://github.com/vllm-project/vllm/blob/v0.11.0rc1/examples/offline_inference/prompt_embed_inference.py

[hwj-cl]

Your current environment

A2服务器上

🐛 Describe the bug

0.11.0rc1镜像：跑prompt_embed_inference.py脚本时，在最开始初始化的时候卡住，查看日志没有报错，NPU显存占用117左右；

0.9.1镜像：跑prompt_embed_inference.py通过，有类似以下输出
[Single Inference Output]

The capital of France is Paris. It is located in the northwest of the country

[Batch Inference Outputs]

Q1:Please tell me about the capital of France,
Al:The capital of France is Paris,It is located in the 1le-de-F

A2: The day with the longest duration in a year is typically the summer solstice,
Q2:When is the day longest during the year?

Q3:Where is bigger, the moon or the sum?
A3: The sun is significantly larger than the moon Here are some key facts to compare

[zhangxinyuehfad]

You can set export VLLM_WORKER_MULTIPROC_METHOD=spawn , and use newer code (#3026 supported prompt embeddings ).

ncrease the number of supported shapes, set HCCL_OP_EXPANSION_MODE=AIV.
(EngineCore_DP0 pid=18193) INFO 11-14 09:51:52 [utils.py:523] No adjustment needed for ACL graph batch sizes: LlamaForCausalLM model (layers: 16) with 67 sizes
(EngineCore_DP0 pid=18193) WARNING 11-14 09:51:52 [platform.py:324] If chunked prefill or prefix caching is enabled, block size must be set to 128.
INFO 11-14 09:51:53 [llm.py:306] Supported_tasks: ['generate']
Downloading Model from https://www.modelscope.cn to directory: /root/.cache/modelscope/hub/models/LLM-Research/Llama-3.2-1B-Instruct
2025-11-14 09:51:54,282 - modelscope - INFO - Target directory already exists, skipping creation.
WARNING 11-14 09:51:54 [model.py:1389] Default sampling parameters have been overridden by the model's Hugging Face generation config recommended from the model creator. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
Adding requests: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 478.15it/s]
Processed prompts:   0%|                                                                                                                                                                                                | 0/1 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s](EngineCore_DP0 pid=18193) INFO 11-14 09:51:54 [acl_graph.py:188] Replaying aclgraph
Processed prompts: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  4.28it/s, est. speed input: 188.80 toks/s, output: 68.65 toks/s]

[Single Inference Output]
------------------------------
The capital of France is Paris. It is located in the Île-de-F
------------------------------
Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3821.11it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 17.41it/s, est. speed input: 772.73 toks/s, output: 278.86 toks/s]

[Batch Inference Outputs]
------------------------------
Q1: Please tell me about the capital of France.
A1: The capital of France is Paris. It's the most populous city in France and

Q2: When is the day longest during the year?
A2: The day with the longest duration during the year is typically the summer solstice,

Q3: Where is bigger, the moon or the sun?
A3: The sun is significantly larger than the moon. Here are some key facts to put

------------------------------

[wangxiyuan]

@hwj-cl Prompt Embeddings feature doesn't work in v0.11.0 release, and we don't have plan to cherry-pick. You can use main branch or wait for v0.11.1rc1 release.

@zhangxinyuehfad Please update the example to set the VLLM_WORKER_MULTIPROC_METHOD by default.