vllm-ascend-model-adapter

To adapt a Hugging Face model for vLLM Ascend, this Skill applies a structured workflow that validates deterministic startup and inference, delivering a single signed commit for integration.

Validating model inference on Ascend hardware requires stage-based checks using dummy and real-weight gates to verify startup, smoke tests, and feature support across ACLGraph and EP paths.

vLLM Ascend adaptation supports multimodal paths where applicable, running stage-based validations to ensure deterministic startup, inference, and feature compatibility for the target model.

Ensuring deterministic startup for vLLM Ascend deployment involves running a two-stage adaptation workflow from the workspace that executes fixed implementation roots and validates startup gates.

Adapting models for vLLM Ascend requires fixed implementation roots and specific feature support validation for ACLGraph, EP, and MTP, meaning unsupported features will fail the stage-based validation gates.

You do not need real weights initially; vLLM Ascend integration validation uses a two-stage process starting with dummy weight gates before proceeding to real-weight inference checks.