t.py

import torch
import torch._inductor.config as inductor_config

inductor_config.optimize_scatter_upon_const_tensor = False

torch.set_default_device("cuda")

M, N = 1024, 2048
x = torch.randint(0, N, (M,), dtype=torch.int64)

@torch.compile
def f(x):
    y = torch.full([M, N], 3.14, dtype=torch.float)
    y.scatter_(1, x.unsqueeze(1), 2.718)
    return y

f(x)   
print("bye")