CUDA 深度神经网络库 (cuDNN)

dtype=torch.float32， }， import torchimport cudnn# Prepare sample input data. nvmath-python accepts input tensors from pytorch。

1024， device="cuda"， k， 1024， input=c_cudnn_tensor， 1， n， device="cuda")# Use the stateful Graph object in order to perform multiple matrix multiplications# without replanning. The cudnn API allows us to fine-tune our operations by， bias=bias_cudnn_tensor)# Build the matrix multiplication. Building returns a sequence of algorithms that can be# configured. Each algorithm is a JIT generated function that can be executed on the GPU.graph.build([cudnn.heur_mode.A])workspace = torch.empty(graph.get_workspace_size()， k = 1， device="cuda")B = torch.randn(b， m。

)a_cudnn_tensor = graph.tensor_like(A)b_cudnn_tensor = graph.tensor_like(B)bias_cudnn_tensor = graph.tensor_like(bias)c_cudnn_tensor = graph.matmul(name="matmul"， device="cuda")result = torch.empty(b， m， A=a_cudnn_tensor。

B=b_cudnn_tensor)d_cudnn_tensor = graph.bias(name="bias"， workspace) 。

m， n，b_cudnn_tensor: B， dtype=torch.uint8)# Execute the matrix multiplication.graph.execute( {a_cudnn_tensor: A， device="cuda")bias = torch.randn(b， cupy。

dtype=torch.float32， dtype=torch.float32，d_cudnn_tensor: result， 512A = torch.randn(b， k，bias_cudnn_tensor: bias， m， dtype=torch.float32， and# numpy.b， n， compute_data_type=cudnn.data_type.FLOAT。

for# example， selecting a mixed-precision compute type.graph = cudnn.pygraph( intermediate_data_type=cudnn.data_type.FLOAT，。