TinyGrad: AI for everyone
With the goal to enable AI to everyone, tinygrad is the new ML framework which is much smaller and simpler than pyTorch.
According to their website, they have three advantages over pyTorch:
1. It compiles a custom kernel for every operation, allowing extreme shape specialization.
2. All tensors are lazy, so it can aggressively fuse operations.
3. The backend is 10x+ simpler, meaning optimizing one kernel makes everything fast.
stable diffusion and Llama are already implemented in tinygrad.
Currently they are focused on putting AMD at MLperf benchmark which is annually dominated by NVIDIA.
for more posts like this follow Abhiram Kandiyana