Segfault when using addcmul in the frontend
jacobhinkle opened this issue Β· comments
Jacob Hinkle commented
π Describe the bug
The following test segfaults somewhere in nvfuser::python_frontend::OpRecord<nvfuser::TensorView*, nvfuser::TensorView*, nvfuser::TensorView*, nvfuser::TensorView*, nvfuser::TensorView*>::operator==()
:
def test_addcmul(self):
inputs = [
torch.randn(4, device="cuda", dtype=torch.float32),
torch.randn(4, device="cuda", dtype=torch.float32),
torch.randn(4, device="cuda", dtype=torch.float32),
]
def fusion_func(fd: FusionDefinition):
t0 = fd.from_pytorch(inputs[0])
t1 = fd.from_pytorch(inputs[1])
t2 = fd.from_pytorch(inputs[2])
c0 = fd.define_constant(0.1)
t3 = fd.ops.addcmul(t0, t1, t2, c0)
fd.add_output(t3)
nvfout, _ = self.exec_nvfuser(fusion_func, inputs)
torch_out = torch.addcmul(*inputs, 0.1)
self.assertEqual(nvfout[0], torch_out)
Versions
TOT