×

签到

分享到微信

打开微信,使用扫一扫进入页面后,点击右上角菜单,

点击“发送给朋友”或“分享到朋友圈”完成分享

MLU220支持matmul吗 解决中 qiuming2023-10-12 10:40:50 回复 1 查看 技术答疑 使用求助
MLU220支持matmul吗
分享到:

matmul算子转Pytorch是会报错

import torch

from torch import nn

from torch.nn import functional as F

import torch_mlu

import torch_mlu.core.mlu_model as ct

import torch_mlu.core.mlu_quantize as mlu_quantize


class NetWork(nn.Module):

    def __init__(self):

        super(NetWork, self).__init__()

        self.conv2d_50 = nn.Conv2d(stride=[1, 1], padding=[1, 1], dilation=[1, 1], groups=1, bias=False, in_channels=8, out_channels=8, kernel_size=[3, 3])

        self.conv2d_50.weight.data.normal_(0.0,0.02)


        self.relu = nn.ReLU()


    def forward(self, image):

        image = self.conv2d_50(image)

        image = self.relu(image)

        image2 = self.conv2d_50(image)

        image2 = self.relu(image2)

        result = torch.matmul(image, image2)

        return result

        #return image


torch.set_grad_enabled(False)


model = NetWork()

image = torch.randn([1, 8,15,15])


out = model(image)


qconfig = {'iteration': 2, 'firstconv': False, 'mean':[0.485,0.456,0.406], 'std':[0.229,0.224,0.225]}

quantized_model = torch_mlu.core.mlu_quantize.quantize_dynamic_mlu(model, qconfig, dtype='int8', gen_quant=True)

out = quantized_model(image)


torch.save(quantized_model.state_dict(), "test_matmul_quantize.pth")


quantized_model = torch_mlu.core.mlu_quantize.quantize_dynamic_mlu(model)

state_dict = torch.load("test_matmul_quantize.pth")

quantized_model.load_state_dict(state_dict)



ct.set_core_number(4)

ct.set_core_version('MLU220')

ct.set_device(-1)

ct.set_input_format(0)

ct.save_as_cambricon("./test_matmul")

randn_mlu = image

randn_mlu = randn_mlu.to(ct.mlu_device())

quantized_model.to(ct.mlu_device())

net_traced = torch.jit.trace(quantized_model, randn_mlu, check_trace=False)

net_traced(randn_mlu)




###########################################################################

错误信息如下:

(pytorch_mlu) [root@localhost paddle2pytorch]# python test_matmul.py 

CNML: 7.10.2 ba20487

CNRT: 4.10.1 8532541

2023-10-12 10:19:49.359124: [cnrtWarning] [744912] [Card : NONE] Failed to initialize CNDEV. Host manage interface disabled 

2023-10-12 10:19:49.363088: [cnrtError] [744912] [Card : NONE] No MLU can be found !

2023-10-12 10:19:49.363106: [cnrtError] [744912] [Card : NONE] Error occurred in cnrtInit during calling driver interface.

2023-10-12 10:19:49.363117: [cnrtError] [744912] [Card : NONE] Return value is 5, MLU_ERROR_NO_DEVICE, means that "No useful mlu device"

2023-10-12 10:19:49.363141: [cnmlError] No MLU device

2023-10-12 10:19:49.363318: [cnmlError] No MLU device

WARNING:root:The 'iteration' key in qconfig has been deprecated in this version. Users needn't

set this key. We will calculate quantization parameters by real forward times controlled by users.

WARNING:root:

You are using conv, but set mean and std. Mean and std are only used

for first conv, which will not be used in the situation. Please do not set them!

2023-10-12 10:19:49.440627: [cnrtError] [744912] [Card : NONE] input param  is invalid device handle in cnrtSetCurrentDevice

[WARNING][/pytorch/catch/torch_mlu/csrc/jit/passes/segment_graph.cpp][line:41][MLUSupport][thread:140522997507904][process:744912]: 

[Fusion Segment] Please check mlu_functions.yaml && Maybe MLU fusion does NOT supports op: matmul

%31 : Tensor = aten::matmul(%input.3, %image2), scope: NetWork # test_matmul.py:23:0


[WARNING][/pytorch/catch/torch_mlu/csrc/jit/passes/segment_graph.cpp][line:41][MLUSupport][thread:140522997507904][process:744912]: 

[Fusion Segment] Please check mlu_functions.yaml && Maybe MLU fusion does NOT supports op: matmul

%31 : Tensor = aten::matmul(%input.3, %image2), scope: NetWork # test_matmul.py:23:0


[WARNING][/pytorch/catch/torch_mlu/csrc/jit/passes/segment_graph.cpp][line:41][MLUSupport][thread:140522997507904][process:744912]: 

[Fusion Segment] Please check mlu_functions.yaml && Maybe MLU fusion does NOT supports op: matmul

%31 : Tensor = aten::matmul(%input.3, %image2), scope: NetWork # test_matmul.py:23:0


[WARNING][/pytorch/catch/torch_mlu/csrc/jit/passes/segment_graph.cpp][line:41][MLUSupport][thread:140522997507904][process:744912]: 

[Fusion Segment] Please check mlu_functions.yaml && Maybe MLU fusion does NOT supports op: matmul

%31 : Tensor = aten::matmul(%input.3, %image2), scope: NetWork # test_matmul.py:23:0


[WARNING][/pytorch/catch/torch_mlu/csrc/jit/passes/segment_graph.cpp][line:41][MLUSupport][thread:140522997507904][process:744912]: 

[Fusion Segment] Please check mlu_functions.yaml && Maybe MLU fusion does NOT supports op: matmul

%31 : Tensor = aten::matmul(%input.3, %image2), scope: NetWork # test_matmul.py:23:0


[WARNING][/pytorch/catch/torch_mlu/csrc/jit/passes/segment_graph.cpp][line:41][MLUSupport][thread:140522997507904][process:744912]: 

[Fusion Segment] Please check mlu_functions.yaml && Maybe MLU fusion does NOT supports op: matmul

%31 : Tensor = aten::matmul(%input.3, %image2), scope: NetWork # test_matmul.py:23:0


Traceback (most recent call last):

  File "test_matmul.py", line 54, in <module>

    net_traced(randn_mlu)

  File "/mnt/disk1/anaconda3/envs/pytorch_mlu/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in __call__

    result = self.forward(*input, **kwargs)

RuntimeError: isTensor() INTERNAL ASSERT FAILED at /pytorch/aten/src/ATen/core/ivalue_inl.h:86, please report a bug to PyTorch. Expected Tensor but got IntList

The above operation failed in interpreter, with the following stack trace:

test_matmul.py(23): forward

/mnt/disk1/anaconda3/envs/pytorch_mlu/lib/python3.6/site-packages/torch/nn/modules/module.py(525): _slow_forward

/mnt/disk1/anaconda3/envs/pytorch_mlu/lib/python3.6/site-packages/torch/nn/modules/module.py(539): __call__

/mnt/disk1/anaconda3/envs/pytorch_mlu/lib/python3.6/site-packages/torch/jit/__init__.py(997): trace_module

/mnt/disk1/anaconda3/envs/pytorch_mlu/lib/python3.6/site-packages/torch/jit/__init__.py(858): trace

test_matmul.py(53): <module>


版权所有 © 2024 寒武纪 Cambricon.com 备案/许可证号:京ICP备17003415号-1
关闭