starmountain1997

Claude Code skills for Ascend NPU inference, model quantization, LLM evaluation, and structured Git commits

Ascend NPU hardware entry point — health check, environment setup, shell script template. Starting point for any Ascend workflow.

vLLM-Ascend serving toolchain — install, offline validation, scenario tuning, online serving, contribution guide.

Download models from ModelScope or HuggingFace to local storage before inference, quantization, or evaluation.

Model quantization on Ascend NPUs — W4A8/W8A8/W4A4, one-click and custom YAML, MoE mixed precision, VLM support, accuracy recovery.

AISBench evaluation framework — accuracy benchmarks (GSM8K, MMLU, AIME) and performance benchmarks against vLLM services.

Structured Git commits with WHAT/WHY/HOW format, optimized as AI context for future sessions.

MindStudio-Modeling performance evaluation framework for Ascend NPUs. Provides tensor-cast for model-level analysis and serving-cast-simulation for LLM serving parameter tuning.

Verify NPU competition entries — clone work and original model repos, check model class name overlap, validate script syntax, and test runnability.