القائمة
Use when validating repo-owned inference engines or model lanes through tracked TOML scenarios after package changes or host repairs.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Use when Arch package changes need a host handoff, rebuild, publish, install, deploy, reinstall, downgrade, local pacman repo refresh, or post-change package verification in this repo.
Use when updating, auditing, or extending this repo's package set, checking the 24-hour dependency freshness sweep, or reconciling upstream, Arch, AUR, CachyOS, or Blackcat Informatics recipe drift.
Use when ROCm, vLLM, AITER, FlashAttention, MIGraphX, quantization, profiling, or inference-optimization references need to become package notes, backlog items, scenarios, or troubleshooting docs in this repo.
| name | run-local-inference-scenarios |
| description | Use when validating repo-owned inference engines or model lanes through tracked TOML scenarios after package changes or host repairs. |
Use python tools/run_inference_scenarios.py for repo-owned inference
validation. The harness is serial, scenario-driven, and logs to a predictable
run root.
When a scenario comes from ROCm inference optimization, quantization,
profiling, MIGraphX, Torch-MIGraphX, FlashAttention, AITER, Triton, or
Composable Kernel references, open
docs/maintainers/rocm-inference-reference.md before creating TOML coverage.
python tools/run_inference_scenarios.py \
--scenario vllm.gemma4.26b-a4b.text.basic \
--scenario vllm.gemma4.26b-a4b.server.basic \
--model-path google/gemma-4-26B-A4B-it=/absolute/path/to/google/gemma-4-26B-A4B-it
python tools/run_inference_scenarios.py --engine vllm
python tools/run_inference_scenarios.py --engine lemonade
If no selector args are given, the tool prompts on a TTY and fails fast otherwise.
Run live vLLM, PyTorch HIP, ROCm, rocminfo, amd-smi, and server-smoke
scenarios outside the Codex bwrap sandbox. In Codex, request
sandbox_permissions=require_escalated for commands that need to open
/dev/kfd or /dev/dri/renderD*; adding those paths as writable roots is not
enough proof of usable device access.
The sandbox is fine for unit tests, catalog tests, adapter tests, and runner dry-runs that do not initialize ROCm. Do not classify a live vLLM scenario as GPU-unavailable until the command has been retried with host device access.
docs/worklog/inference-runs/<timestamp>/.plan.json, result.json, stdout.log,
stderr.log, and server.log when applicable.amd-smi process tables and fail early on a
preexisting stale VLLM::EngineCore.tools/amerge run ... first.