Problems with clamp/clamp_min/clamp_max/minimum/maximum #93784
Description
Repro:
import torch
import torchdynamo
import torchinductor
from torchinductor import config
torchinductor.config.debug = True
def relu(x0):
return torch.nn.functional.relu(x0)
def maximum(x0, x1):
return (torch.clamp_min(x0, 0.5), torch.maximum(x0, x1))
def clamp_min_tensor(x0, x1):
return torch.clamp_min(x0, x1)
device='cuda'
dtype=torch.half
x0 = torch.arange(-3, 4, 1, device=device, dtype=dtype)
x1 = torch.zeros_like(x0)
x0[1]=float('nan')
x0[-1]=float('nan')
x1[2]=float('nan')
print(x0)
optimize_ctx = torchdynamo.optimize("inductor")
with optimize_ctx:
out_inductor = (relu(x0), maximum(x0, x1))
out_eager = (relu(x0), maximum(x0, x1))
print(out_inductor) #clamp_min doesn't propagate nans, maximum propagates nans only from one of the args
print(out_eager)
x0=torch.randint(4,(7,), device=device)
with optimize_ctx:
out_inductor=maximum(x0, x1)
out_eager=maximum(x0, x1)
print(out_inductor) #clamp_min doesn't type promote
print(out_eager)
x0 = torch.randn(7, device=device, dtype=dtype)
with optimize_ctx:
out_inductor=clamp_min_tensor(x0, x1) #errors out
- clamp_min doesn't propagate nans,
- maximum propagates nans only from the second arg
- clamp_min doesn't type promote
- clamp_min with tensor arg errors
relu propagates nans correctly, luckily
More systematically, we should be running OpInfo tests with inductor.