import torch

def fn(a, b):
    return torch._foreach_sub(a, b, alpha=2.0)

torch.manual_seed(42)
a = [torch.randn(4, device='cuda') for _ in range(3)]
b = [torch.randn(4, device='cuda') for _ in range(3)]

eager = fn(a, b)
compiled = torch.compile(fn)([t.clone() for t in a], [t.clone() for t in b])

for i in range(3):
    print(f"eager[{i}]:    {eager[i].tolist()}")
    print(f"compiled[{i}]: {compiled[i].tolist()}")
    print(f"  match: {torch.equal(eager[i], compiled[i])}")  # False

# compiled output exactly matches alpha=1 (no alpha):
no_alpha = torch._foreach_sub(a, b)
for i in range(3):
    assert torch.equal(compiled[i], no_alpha[i]), "compiled should match no-alpha"
print("\nalpha was silently ignored: compiled == no_alpha for all tensors")

---

eager[0]:    [1.2315161228179932, -0.292142391204834, -1.423011302947998, 2.6724295616149902]
compiled[0]: [0.7127674221992493, 0.9346156120300293, -0.7975307703018188, 1.7607448101043701]
  match: False
...
alpha was silently ignored: compiled == no_alpha for all tensors

---

# line 7953 — _foreach_add correctly passes allow_alpha:
foreach_add_list = register_foreach_pointwise(
    aten._foreach_add.List, add, allow_alpha=True
)

# line 7963 — _foreach_sub is missing allow_alpha:
register_foreach_pointwise(aten._foreach_sub.List, sub)  # ← alpha silently dropped

---

register_foreach_pointwise(aten._foreach_sub.List, sub, allow_alpha=True)

---

PyTorch: 2.13.0.dev20260513+cu126
GPU: NVIDIA Tesla T4
CUDA: 12.6
Triton: 3.3.1