[NPU] Adapt vLLM-Ascend 0.23 runtime

[addsubmuldiv]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Adapt the NPU runtime integration to vLLM-Ascend 0.23.0.

Changes

• Register the custom OPP path bundled with the vllm-ascend wheel before torch_npu initializes CANN.
• Detect and use the native vLLM-Ascend fused-MoE LoRA implementation when available.
• Skip legacy packed-LoRA expansion patches when the installed vLLM version already supports them natively.
• Update Chinese and English NPU documentation with the vLLM-Ascend 0.23.0 installation instructions:
  • Python 3.12
  • PyTorch 2.10.0 / TorchNPU 2.10.0.post4
  • vLLM 0.23.0 / vLLM-Ascend 0.23.0
  • Required Ascend PyPI mirror indexes

Compatibility

Please install CANN, PyTorch, TorchNPU, and vLLM-Ascend according to the official compatibility matrix.

Experiment results

• NPU environment initialization
• vLLM-Ascend inference
• MoE LoRA rollout/training

[Copilot]

Pull request overview

Adapts the NPU integration and documentation for vLLM-Ascend 0.23.0.

Changes:

• Registers wheel-bundled custom OPPs before TorchNPU initialization.
• Detects native fused-MoE and packed-LoRA capabilities.
• Updates Chinese and English installation instructions.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
swift/model/npu_patch/vllm_ascend_lora.py	Adds native LoRA capability detection.
swift/model/npu_patch/env.py	Bootstraps the custom OPP environment.
docs/source/BestPractices/NPU-support.md	Updates Chinese NPU installation guidance.
docs/source_en/BestPractices/NPU-support.md	Updates English NPU installation guidance.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[hjh0119]

Could you confirm whether the earlier vllm_ascend patch can be safely removed starting from version 0.23?

[addsubmuldiv]

Could you confirm whether the earlier vllm_ascend patch can be safely removed starting from version 0.23?

For now, could we keep compatibility with vLLM-Ascend 0.18 as well? Removing the legacy patch would require additional validation to ensure there are no regressions for existing NPU users. We can handle the cleanup in a follow-up PR once that verification is complete.