Skill

kaggle-llm-serving

Use when an LLM inference notebook on Kaggle GPUs (4x L4, 2x T4, single 24 GB cards) is throughput-bound — tokens/s too low for the runtime limit, choosing tensor vs data parallel, fitting a 20-30B model on one 24 GB card, KV-cache or Mamba/GDN state limiting concurrency, MTP/speculative decoding depth, int4/fp8 weights or KV, vLLM patch stacks (HyperQwen, Merlin), or benchmarking serving configs on RunPod before spending Kaggle quota

Bundled with kaggle-tools. Its trust score, security findings and other files are on the plugin page.

npx [email protected] add yassineAlouini/kaggle-tools