Salman Mohammadi

Thanks Lewis!

Currently we only support Python. However, there are WebAssembly runtimes for many other languages, so we think this approach could work for other languages. If there's community interest here I could definitely add support for this.

For benchmarking, see this script https://gist.github.com/SalmanMohammadi/ea25d12340851ce033885564dc4bc720
You should be able to run this after installing the grpo_code repo using MAX_WORKERS={max_workers} test_grpo_mp.py.

For reference, with my meagre 10 processing cores on my laptop, I get:

The processing speedups scale well with num_generations * batch_size * n_test_cases - for this blogpost we found MAX_WORKERS=64 to provide ~10x speedups when you have the processor count available in high-compute scenarios.