Relay failed to build models exported from pytorch

Questions
#1

Hi there,

While I was following the tutorial relay quick start, I tried to load a module from pytorch but it raises segmentation fault error. The TVM I am using the latest commit bff98843bef9a312587aaff51b679d9b69a7d5a7 and the code to reproduce is attached below

from tvm import relay
import tvm
import numpy as np

import torch
import torch as th
import torch.nn as nn
from torchvision import models

import torch.onnx 
from tvm import relay, auto_scheduler

model = nn.Sequential(
    nn.Conv2d(3, 3, kernel_size=3, padding=1),
    nn.BatchNorm2d(3),
    # nn.Dropout()
)

input_shape = [1, 3, 224, 224]
input_data = torch.randn(input_shape)
scripted_model = torch.jit.trace(model, input_data).eval()
input_name = "input0"
shape_list = [(input_name, input_data.shape)]
mod, params = relay.frontend.from_pytorch(scripted_model, shape_list)
print(mod['main'])

opt_level = 3
# target = tvm.target.cuda()
target = "llvm"
with tvm.transform.PassContext(opt_level=opt_level):
    lib = relay.build(mod, target=target, params=params)
#2

My environemnt

ubuntu 20.04 | gcc: 9.3 | llvm: 10.0 | nvcc: 11.1

#3

Hi @Lyken17,

I also ran into this issue recently. It turned out to be conflicting symbols between PyTorch and TVM, see https://github.com/apache/tvm/issues/9362#issuecomment-955263494 for the resolution. Alternatively, a quicker (but less elegant) solution is to import torch before tvm. Hope this helps!

1 Like
#4

Notice one very interesting one that might be helpful for this bug

I tried to compile the models exported from mxnet

from tvm import relay
import tvm
import numpy as np

import torch
import torch as th
import torch.nn as nn
from torchvision import models
import torch.onnx 

from tvm import relay, auto_scheduler
from tvm.relay import testing


mod, params = relay.testing.resnet.get_workload(
    num_layers=18, batch_size=1, image_shape=(3, 224, 224)
)

opt_level = 3
# target = tvm.target.cuda()
target = "llvm"
with tvm.transform.PassContext(opt_level=opt_level):
    lib = relay.build(mod, target=target, params=params)

print("build sucessful")

and it also raises the same segment fault error. However, if I comment all pytorch imports

import torch
import torch as th
import torch.nn as nn
from torchvision import models
import torch.onnx

then the relay.build works without problem. Any thoughts on the strange behavior?

#5

Oh! Thanks for the information. You saved my day.