I don’t have deep knowledge of ML, but if I had to recommend one, it would be PyTorch. PyTorch models can be converted to TensorRT format via ONNX, making them suitable for everything from research purposes to execution on edge devices. Even considering the hellishly large library size and nightmarish installation experience due to CUDA, I think it’s an excellent library.
The main libraries on Hugging Face are PyTorch wrappers (inheriting from torch.nn.Module). This means that if you learn PyTorch, you can use most of the model weights and code available on Hugging Face as learning materials. This is a significant advantage.