quark-torch-llm-eval

name

quark-torch-llm-eval

description

End-to-end LLM accuracy evaluation on AMD ROCm (ROCm-only) — docker container setup OR host (no-docker) runtime, vLLM/SGLang/ATOM serving, lm-eval / lighteval / evalscope benchmarks. Use when the user wants to evaluate, benchmark, or compare an LLM's accuracy. Trigger for "evaluate this model", "run gsm8k/mmlu/mmlu_pro/aime/gpqa/hellaswag/arc", "test accuracy", "measure perplexity", "compare quantized model accuracy", "does this mxfp4 model lose accuracy". For evaluating Quark Agent Skills themselves, use quark-torch-eval-runner instead.

Read and follow the instructions in .claude/skills-impl/l1-atomic/torch/quark-torch-llm-eval/SKILL.md.