vllm-ascend-workspace: Skill for Codex

vLLM Ascend Correctness Validation

Produce traceable correctness evidence instead of treating a successful request as proof of correctness.

Workflow

1. Define cases and the smallest affected validation matrix.
2. Create independent baseline and candidate code states or eager and graph states.
3. Before remote execution, use remote-code-parity for each state.
4. For offline cases, run scripts/remote_correctness_harness.py inside the remote NPU container. The harness prioritizes the materialized vllm/ and vllm-ascend/ source roots and propagates them through PYTHONPATH, so launching from /vllm-workspace or spawning a python -m child cannot resolve the outer repository directory as a false vllm namespace package.
5. For online cases, use vllm-ascend-serving, then run the same harness in online-chat mode.
6. Use scripts/correctness_run.py init to create the run directory and Run Manifest v1.
7. Use scripts/correctness_run.py compare to normalize the evidence into a classification and report.
8. Route failures:
   • eager pass and graph fail: vllm-ascend-graph-debug;
   • multi-rank hang or inconsistent rank metadata: vllm-ascend-distributed-debug when available;
   • performance-only change: vllm-ascend-performance-regression when available;
   • task metric execution: use the bundled AISBench adapter.

Do not run the full execution-mode, parallelism, and feature Cartesian product. Select cases from the code impact and .agents/knowledge/validation-rules.yaml; record omitted combinations as risks.

Determinism

For exact token or text comparison:

• set temperature=0;
• set an explicit seed;
• keep prompts, messages, chat template, max tokens, model weights, tokenizer, parallel topology, and feature flags identical;
• repeat each case at least twice when nondeterminism is suspected;
• never compare baseline and candidate results produced from different case files.

Use task-level metrics when exact output is not an appropriate acceptance condition. Use numeric tolerances only for explicitly captured logits, hidden states, KV samples, or other numeric evidence.