Closed
Description
Supplying two integer inputs results in a type promotion to float which does not match C-Style div.
Repro:
import torch
from nvfuser import FusionDefinition, DataType
inputs = [
torch.randint(255, (4, 4), device='cuda'),
torch.randint(255, (4, 4), device='cuda'),
]
def fusion(fd: FusionDefinition):
t0 = fd.from_pytorch(inputs[0])
t1 = fd.from_pytorch(inputs[1])
t2 = fd.ops.div(t0, t1)
fd.add_output(t2)
with FusionDefinition() as fd:
fusion(fd)
out = fd.execute(inputs)
print(inputs[0])
print(inputs[1])
print(out)
Output:
tensor([[ 68, 177, 192, 186],
[186, 42, 199, 181],
[ 56, 103, 79, 31],
[199, 69, 54, 241]], device='cuda:0')
tensor([[ 98, 149, 45, 103],
[ 90, 86, 58, 195],
[ 77, 211, 142, 85],
[ 11, 225, 232, 241]], device='cuda:0')
[tensor([[ 0.6939, 1.1879, 4.2667, 1.8058],
[ 2.0667, 0.4884, 3.4310, 0.9282],
[ 0.7273, 0.4882, 0.5563, 0.3647],
[18.0909, 0.3067, 0.2328, 1.0000]], device='cuda:0')]
Metadata
Metadata
Assignees
Labels
No labels