🤖 TPU Model Performance Auto-optimization

🤖 TPU Model Performance Auto-optimization

I’d like to make a very bold statement: given a sufficiently capable LLM, the right profiling tools, and a knowledge base that includes the model’s + framework’s source, an autonomous agent can drive any (model, hardware) pair to state-of-the-art performance for that combination.

If you think about this, conceptually same applicable to the engineers (can extrapolate to a new-grad without practical experience) we need tools (xprof), knowledge base (TPU optimization information), code base to work with, and reference optimized code base as bonus to be be able optimize models.

Anthropic recently published an article on recursive self improvment, where the claim is that: In the future, agents could become capable enough to build and train models themselves: https://www.anthropic.com/institute/recursive-self-improvement

ho

For performance model optimization, which is a different but, but nevertherless extremely complex and depth domain it is already possible, at least partially.

Here is a repo that proves and demonstrates that all of that is possible already: https://github.com/vlasenkoalexey/tpu_performance_autoresearch_wiki