Q4 2025 VERL+NPU Key Priorities

[zheliuyu]

Continued Work: Q3 Incomplete Tasks

Q3 roadmap: #2171

New features in Q4

• Update the versions of torch_npu, cann and vllm-ascend.
• Retool RL.
• verl+SGLang on npu? Yes! @lbk-sys, [Tracking]feat: support verl + SGLang on npu. #2916

[zheliuyu]

@FightingZhen cc

[FightingZhen]

Torch.compile maybe still can not be supported on Ascend NPU yet, because next stable vllm-ascend version will be 0.9.1, which will strictly install torch_npu==2.5.1. Torch.compile can be used with torch_npu >=2.6.x.

[tardis-key]

May I ask where this conclusion comes from? I notice that the current torch version requirements on 091dev are PyTorch >= 2.5.1 and torch-npu >= 2.5.1.post1. https://github.com/vllm-project/vllm-ascend/blob/v0.9.1-dev/README.md

[zheliuyu]

May I ask where this conclusion comes from? I notice that the current torch version requirements on 091dev are PyTorch >= 2.5.1 and torch-npu >= 2.5.1.post1. https://github.com/vllm-project/vllm-ascend/blob/v0.9.1-dev/README.md

In practice, the installed version of torch-npu is 2.5.1.post1. https://github.com/vllm-project/vllm-ascend/blob/v0.9.1/pyproject.toml#L15

[zheliuyu]

follow: vllm-project/vllm-ascend#2838

[FightingZhen]

@zheliuyu Add one, support dynamic turn on/off virtual memory switches between training and inference stages