class matmul_transpose (torch.autograd.Function):
使用NPU自定义算子替换原生写法,以提高性能。
Tensor - 输出张量
在动态shape场景中,由于算子限制,不支持Box transformation deltas。
from torch_npu.contrib.function import matmul_transpose output = matmul_transpose(tensor1, tensor2)
>>> tensor1 = torch.randn(68, 5, 75, 16).npu()
>>> tensor1.requires_grad_(True)
>>> tensor2 = torch.randn(68, 5, 75, 16).npu()
>>> tensor2.requires_grad_(True)
>>> output = matmul_transpose(tensor1, tensor2)
>>> output.sum().backward()