We fine-tuned Llama 405B on AMD GPUs

publish.obsidian.md

495 points

2 years ago

Hey HN, we recently fine-tuned the llama3.1 405B model on 8xAMD MI300x GPUs using JAX instead of PyTorch. JAX's advanced sharding APIs allowed us to achieve great performance. Check out our blog post to learn about the cool sharding tricks we used. We've also open-sourced the code: https://github.com/felafax/felafax

We're a small startup building AI infra for fine-tuning and serving LLMs on non-NVIDIA hardware (TPUs, AMD, Trainium).

Problem: Many companies are trying to get PyTorch working on AMD GPUs, but we believe this is a treacherous path. PyTorch is deeply intertwined with the NVIDIA ecosystem in a lot of ways (e.g., `torch.cuda` or scaled_dot_product_attention is an NVIDIA CUDA kernel exposed as a PyTorch function). So, to get PyTorch code running on non-NVIDIA hardware, there's a lot of "de-NVIDIAfying" that needs to be done.

Solution: We believe JAX is a better fit for non-NVIDIA hardware. In JAX, ML model code compiles to hardware-independent HLO graphs, which are then optimized by the XLA compiler before hardware-specific optimization. This clean separation allowed us to run the same LLaMA3 JAX code both on Google TPUs and AMD GPUs with no changes.

Our strategy as a company is to invest upfront in porting models to JAX, then leverage its framework and XLA kernels to extract maximum performance from non-NVIDIA backends. This is why we first ported Llama 3.1 from PyTorch to JAX, and now the same JAX model works great on TPUs and runs perfectly on AMD GPUs.

We'd love to hear your thoughts on our vision and repo!

101 comments