# # PyTorch with CUDA GPU support # # This example shows how you can use CUDA GPUs in Modal, with a minimal PyTorch # image. You can specify GPU requirements in the `stub.function` decorator. import time import modal stub = modal.Stub( "example-import-torch", image=modal.Image.debian_slim().pip_install( "torch", find_links="https://download.pytorch.org/whl/cu116" ), ) @stub.function(gpu="any") def gpu_function(): import subprocess import torch subprocess.run(["nvidia-smi"]) print("Torch version:", torch.__version__) print("CUDA available:", torch.cuda.is_available()) print("CUDA device count:", torch.cuda.device_count()) if __name__ == "__main__": t0 = time.time() with stub.run(): gpu_function.call() print("Full time spent:", time.time() - t0)