What problem does it solve? Running local language models across one laptop or a fleet of machines involves many error-prone steps: checking hardware capacity, finding and downloading the right GGUF weights, sizing context and concurrency, starting the engine, and verifying it actually answers. This Skill guides an agent through that entire lifecycle safely, avoiding common failures like oversized context windows, missing vision projectors, and unverified deployments. ## Core Features & Use Cases - Guided model selection: Asks three plain-language questions (use case, context length, concurrency), then matches answers against models already on disk, the Grid catalog, or Hugging Face GGUF repos. - Safe deployment workflow: Checks host memory via device-info, verifies the llama.cpp binary directly, sizes context against usable bytes, handles vision projector files, and starts models with join using explicit --ctx-size and --max-concurrency. - Verified serving: Waits for relay registration, then proves the model answers with a single bounded max_tokens request before reporting success. - Fleet operations: Connects and selects grids, manages machines over Harness or SSH transports, changes or stops running models, and uses routing, usage, media, and training commands. - Use Case: A user asks to run a coding model on their GPU workstation. The Skill checks the workstation's free memory, finds a suitable GGUF on Hugging Face, pulls it, starts it with 128K context and 2 slots, verifies it responds, and tells the user to pick it from the model dropdown. ## Quick Start Ask the agent to start a coding model with long context on your GPU machine and it will check capacity, pick a fitting model, deploy it, and verify it answers.
