collect-node-data

GPU kernel autotuning data collection is automated by provisioning remote servers, running ε-greedy golden kernel tuning, and merging the results into a local cross-hardware database to improve kernel selection priors.

Cross-hardware autotuning prevents data collisions by using GPU-keyed node storage, ensuring performance measurements from different GPU models never overwrite each other during local database merge operations.

ε-greedy tuning should be used for LLM compiler kernel optimization to avoid prior censoring, ensuring that the autotuning search tree explores sufficiently and does not bias kernel selection priors prematurely.

You can add performance measurements for a new GPU model by renting the specified hardware remotely, setting up the emmy compiler, executing tuning runs, and safely merging the new node data locally.

Node data collection does not overwrite existing hardware measurements because safe merge operations restrict updates to the GPU-keyed node table only, preserving all other hardware data in the shared database.

The best way to expand a cross-hardware autotune dataset is to automate end-to-end remote GPU tuning, which handles server provisioning, compiler setup, and collision-safe data merging without manual intervention.