🐛 Describe the bug
Is this a bug in the nvFuser or the code below is invalid? I think the code is valid, I translated the trace of torch._refs.add to nvFuser Python API calls. There's no error with ATen execution.
import torch
from torch._C._nvfuser import Fusion, FusionDefinition
# Construct and Define Fusion
fusion = Fusion()
with FusionDefinition(fusion) as fd :
t0 = fd.define_tensor(2)
t1 = fd.define_tensor(1)
fd.add_input(t0)
fd.add_input(t1)
t0_b = fd.Ops.broadcast_in_dim(t0, [3, 3], [0, 1])
t1_b = fd.Ops.broadcast_in_dim(t1, [3, 3], [1])
t2 = fd.Ops.add(t0_b, t1_b)
fd.add_output(t2)
fusion.print_ir()
# Execute Fusion
input1 = torch.ones(3, 1, device='cuda')
input2 = torch.ones(3, device='cuda')
fusion.execute([input1, input2])fusion.execute call raises
RuntimeError: Attempting to bind T0.size[1] to 3but it's already set to 1
Versions
.
🐛 Describe the bug
Is this a bug in the nvFuser or the code below is invalid? I think the code is valid, I translated the trace of
torch._refs.addto nvFuser Python API calls. There's no error with ATen execution.fusion.executecall raisesVersions
.