collect-node-data

What problem does it solve?

Manually collecting per-GPU kernel performance node data for LLM compiler autotuning requires remote server management, careful data merging to avoid hardware collisions, and time-consuming tuning runs, which is error-prone and inefficient for teams expanding their cross-hardware autotune datasets.

Core Features & Use Cases

• End-to-end remote tuning automation: Rents a single specified GPU, sets up the emmy compiler, runs ε-greedy golden kernel tuning to avoid prior censoring, and merges results into the local cross-hardware autotune database.
• Collision-safe data storage: All node data is GPU-keyed, so measurements from different cards never overwrite each other in the shared local database.
• Use Case: If your team needs to improve kernel selection priors for a new H200 GPU, use this skill to automatically collect the required node data without manually provisioning servers or handling data merges.

Quick Start

Use the collect-node-data skill to gather node data for an H200 GPU and merge it into your local autotune database.