search

airockchip / rknn-toolkit2

Other
729 stars 77 forks source link

The following operators are not implemented: ['aten::scaled_dot_product_attention'] #27

Open stevenlafl opened 4 months ago

stevenlafl commented 4 months ago

I'm attempting to convert a large (really large, 26GB model) to RKNN and I came across this:

E load_pytorch: Traceback (most recent call last):
E load_pytorch:   File "rknn/api/rknn_base.py", line 1616, in rknn.api.rknn_base.RKNNBase.load_pytorch
E load_pytorch:   File "rknn/base/convertor/torch2onnx/pytorch.py", line 4574, in rknn.base.convertor.torch2onnx.pytorch.from_pytorch
E load_pytorch:   File "rknn/base/convertor/torch2onnx/pytorch.py", line 3657, in rknn.base.convertor.torch2onnx.pytorch.PyTorchOpConverter.report_missing_conversion
E load_pytorch:   File "rknn/api/rknn_log.py", line 92, in rknn.api.rknn_log.RKNNLog.e
E load_pytorch: ValueError: The following operators are not implemented: ['aten::scaled_dot_product_attention']

Is this supported by the RK3588? Will it be supported by the toolkit?

yuyun2000 commented 4 months ago

不支持这个算子的，你可以把他转成onnx然后再转rknn试试