TVMError: Check failed: gv.defined() && gv->IsInstance<GlobalVarNode>():

Troubleshooting
#1

Bug description:
when I use compile a keras model, It crashed. The crash message is shown as followsing:

image
image1181×319 17.9 KB

Conditions(triggering this bug)

  1. target= ‘cuda’
  2. opt_level = 1
  3. use relay.vm.compile() to compile rather than relay.build_module.build()

Only when these three conditions are met at the same time, the bug will be triggered.

The Reproducible script

import keras
import tvm
import tvm.relay as relay

model_path = './xception-imagenet_origin.h5'

predict_model = keras.models.load_model(model_path)
print(predict_model.input)
shape_dict = {'input_7': (1, 3, 299, 299)}
irmod, params = relay.frontend.from_keras(predict_model, shape_dict)
target = 'cuda'
ctx = tvm.gpu(0)

with tvm.transform.PassContext(opt_level=1):
   # graph, lib, params = relay.build_module.build(irmod, target, params=params)
   vm_exec = relay.vm.compile(irmod, target, params=params)

Environment

TVM : 0.8.dev0
OS: Ubuntu 16.04

Anyone can download this Keras model from the below link:

#2

@FrozenGene @merrymercy @comaniac Could you give me some advice?