Use this skill for Modal job orchestration with `.map`, `.starmap`, `.spawn`, `.spawn_map`, or `@modal.batched`. Trigger when the user needs to fan out work across Modal containers, collect results in-process, return a pollable job ID from a web server, cap concurrency, recover from partial failures, or batch many small requests into fewer GPU calls. Do not use it for vLLM or SGLang serving, model fine-tuning, or sandbox lifecycle questions.
Use this skill for Modal GPU fine-tuning and post-training: LLM SFT, LoRA, QLoRA, diffusion LoRA, YOLO-style vision training, and GRPO. Trigger when the user wants to adapt model weights, resume or checkpoint training, choose GPUs, stage datasets, or persist training artifacts on Modal. Do not use it for inference serving, detached job orchestration, sandbox lifecycle, or from-scratch pretraining.
Use this skill for self-hosted open-weight text LLM serving on Modal with vLLM or SGLang. Trigger when the user wants an OpenAI-compatible endpoint for a model they host, needs to tune cold starts, latency, concurrency, tensor parallelism, or throughput for text generation, or wants offline batch text inference with the vLLM engine. Do not use it for embeddings, generic Hugging Face pipeline serving, diffusion, fine-tuning, RAG, or apps that only call OpenAI's hosted API.
Use this skill for Modal Sandbox lifecycle and interactive isolated execution. Trigger when the user needs to run untrusted or user-supplied code in a sandbox, keep a long-lived controller process alive, stream stdout from `exec()` calls, reattach to an existing sandbox, upload files at runtime, expose a tunneled HTTP or TCP service, or checkpoint and restore sandbox state. Do not use it for regular `@app.function` deployments, LLM serving, or training workflows.