external-gitcode-ascend-vllm-tests-failure-analysis

Installation
SKILL.md

Upstream vLLM Test Analysis & Adaptation for Ascend NPU

This skill covers two tightly linked workflows for the vllm-ascend project:

  1. Failure Analysis — Diagnose why an upstream vLLM test fails on Ascend NPU, classify the root cause, and determine whether a fix is feasible.
  2. Test Adaptation — Copy an upstream vLLM test into vllm-ascend, adapt it to NPU conventions, debug it until it passes (or conclusively prove it cannot), and make it CI-ready.

Both workflows share the same environment setup and root-cause methodology. The key constraint throughout is: never modify upstream vLLM code — only the test code (once copied into vllm-ascend) and vllm-ascend plugin code may be changed.

1. Check Existing Analysis First

Before analyzing any test, check whether prior work already covers it — this avoids duplication and ensures consistency.

Decision tree:

  1. Read references/ASCEND_ALL_128_TEST_ANALYSIS.md — the consolidated summary table covering 128 tests with root cause, CI verdict, and "should Ascend pass" classification. If the target test appears here and the vllm/vllm-ascend versions haven't changed significantly since the analysis date (2026-03-27), use the existing conclusion. Only re-analyze if the user explicitly requests a fresh run or versions have changed.

  2. If the test appears in references/TEST_FILES_NEED_ANALYSIS.md but NOT in the summary table, it is a known target that has not yet been analyzed. Proceed to full analysis.

Installs
3
GitHub Stars
93
First Seen
May 18, 2026
external-gitcode-ascend-vllm-tests-failure-analysis — ascend-ai-coding/awesome-ascend-skills