onnx libtorch python runtime optim

游雁

2023-04-12 60d38fa9cac0630a82c4238b746e6a7039f848bf

 funasr/export/models/modules/multihead_att.py

@@ -75,8 +75,10 @@
    return x, cache


import torch.fx
torch.fx.wrap('preprocess_for_attn')
torch_version = tuple([int(i) for i in torch.__version__.split(".")[:2]])
if torch_version >= (1, 8):
    import torch.fx
    torch.fx.wrap('preprocess_for_attn')


class MultiHeadedAttentionSANMDecoder(nn.Module):