I actually think the BEAM is an ideal environment for machine learning, sharding things across machines. The only thing I’m not sure of is if PyTorch etc. are more optimised than XLA the backend Axon uses… would be good to see some performance comparisons of a big LLM running on both. For everything else I’d suggest Elixir was a better experience.
Comments
I actually think the BEAM is an ideal environment for machine learning, sharding things across machines. The only thing I’m not sure of is if PyTorch etc. are more optimised than XLA the backend Axon uses… would be good to see some performance comparisons of a big LLM running on both. For everything else I’d suggest Elixir was a better experience.
XLA tends tends to be better optimized for TPUs, Pytorch is better with GPUs, but I believe you can choose a backend when using Nx.