// Adapt and debug existing or new models for vLLM on Ascend NPU. Implement in /vllm-workspace/vllm and /vllm-workspace/vllm-ascend, validate via direct vllm serve from /workspace, and deliver one signed commit in the current repo.
| name | vllm-ascend-model-adapter |
| description | Adapt and debug existing or new models for vLLM on Ascend NPU. Implement in /vllm-workspace/vllm and /vllm-workspace/vllm-ascend, validate via direct vllm serve from /workspace, and deliver one signed commit in the current repo. |
Adapt Hugging Face or local models to run on vllm-ascend with minimal changes, deterministic validation, and single-commit delivery. This skill is for both already-supported models and new architectures not yet registered in vLLM.
references/workflow-checklist.md.references/multimodal-ep-aclgraph-lessons.md (feature-first checklist).references/troubleshooting.md.references/fp8-on-npu-lessons.md.references/deliverables.md.transformers./vllm-workspace/vllm/vllm-workspace/vllm-ascendvllm serve from /workspace with direct command by default.8000 unless user explicitly asks otherwise.--enable-expert-parallel and flashcomm1 checks are MoE-only; for non-MoE models mark as not-applicable with evidence.PYTHONPATH=<modified-src>:$PYTHONPATH unless debugging fallback is strictly needed.git commit -sm ...)./models/<model-name>; if environment differs, confirm with user explicitly)./vllm-workspace/vllm, /vllm-workspace/vllm-ascend)./vllm-workspace/* install.config.json, processor files, modeling files, tokenizer files.vllm/model_executor/models/registry.py.vllm/model_executor/models/;vllm/transformers_utils/processors/ when needed;vllm/model_executor/models/registry.py;/vllm-workspace/vllm-ascend./workspace with --load-format dummy.Application startup complete as pass by itself; request smoke is mandatory./v1/models 200),--load-format dummy and validate with real checkpoint.GET /v1/models first.torch._dynamo + interpolate + NPU contiguous failures on VL paths, try TORCHDYNAMO_DISABLE=1 as diagnostic/stability fallback.skip_tensor_conversion signature mismatch), use text-only isolation (--limit-mm-per-prompt set image/video/audio to 0) to separate processor issues from core weight loading issues.max-model-len=128k + max-num-seqs=16./vllm-workspace/*, backport minimal final diff to current working repo.tests/e2e/models/configs/<ModelName>.yaml following the schema of existing configs (must include model_name, hardware, tasks with accuracy metrics, and num_fewshot). Use accuracy results from evaluation to populate metric values.docs/source/tutorials/models/<ModelName>.md following the standard template (Introduction, Supported Features, Environment Preparation with docker tabs, Deployment with serve script, Functional Verification with curl example, Accuracy Evaluation, Performance). Fill in model-specific details: HF path, hardware requirements, TP size, max-model-len, served-model-name, sample curl, and accuracy table.docs/source/tutorials/models/index.md to include the new tutorial./workspace with direct command.128k + bs16) result is reported, or explicit reason why not feasible.tests/e2e/models/configs/<ModelName>.yaml and follows the established schema (model_name, hardware, tasks, num_fewshot).docs/source/tutorials/models/<ModelName>.md and follows the standard template (Introduction, Supported Features, Environment Preparation, Deployment, Functional Verification, Accuracy Evaluation, Performance).docs/source/tutorials/models/index.md includes the new model entry.Adapt vLLM-Ascend to track upstream vLLM main branch changes incrementally: detect commit drift, plan steps, adapt code, run CI, and commit verified changes. Use whenever the user mentions main2main, upgrading or syncing vllm-ascend to a newer vLLM commit, vLLM API changes breaking vllm-ascend, or provides both a vllm path and vllm-ascend path for syncing. Also triggers on: "vLLM broke our plugin", "bump the vLLM commit", "ascend CI failing after upstream update".
End-to-end release management skill for vLLM Ascend. Creates release checklist issues, identifies critical bugs, runs functional tests, invokes release note generation, and guides through the complete release process.