repro_cache_issue_ws.sh

cat > cache_bc_issue.py << EOF
import torch

torch._inductor.config.autotune_remote_cache = False

def fn(x, y, z):
    return torch.mm((x + y).relu(), z)

x, y, z = [torch.rand((8, 8), device="cuda") for _ in range(3)]

fn_s = torch.compile(fn, mode="max-autotune")

for _ in range(3):
    print(fn_s(x, y, z))
EOF

rm -rf /tmp/torchinductor_$USER
python cache_bc_issue.py
git revert 0861af2596a40fae40adab24131378d4da98addd
python cache_bc_issue.py