In [1]: import warnings
   ...: warnings.simplefilter('error')
   ...: import torch
   ...: print(torch.__version__)
   ...: a, b = torch.rand((2, 32, 32))
   ...: a.requires_grad_()
   ...: optimizer = torch.optim.LBFGS([a])
   ...: loss_fn = lambda x, y: (x-y).pow(2).mean()
   ...: 
   ...: def closure():
   ...:     optimizer.zero_grad()
   ...:     loss = loss_fn(a, b)
   ...:     loss.backward()
   ...:     return loss
   ...: 
   ...: for i in range(100):
   ...:     optimizer.step(closure)
   ...:     print(i, loss_fn(a, b))
   ...: 
2.9.0a0+git23b0334
---------------------------------------------------------------------------
UserWarning                               Traceback (most recent call last)
Cell In[1], line 17
     14     return loss
     16 for i in range(100):
---> 17     optimizer.step(closure)
     18     print(i, loss_fn(a, b))

File ~/code/pytorch/torch/optim/optimizer.py:516, in Optimizer.profile_hook_step.<locals>.wrapper(*args, **kwargs)
    511         else:
    512             raise RuntimeError(
    513                 f"{func} must return None or a tuple of (new_args, new_kwargs), but got {result}."
    514             )
--> 516 out = func(*args, **kwargs)
    517 self._optimizer_step_code()
    519 # call optimizer step post hooks

File ~/code/pytorch/torch/utils/_contextlib.py:120, in context_decorator.<locals>.decorate_context(*args, **kwargs)
    117 @functools.wraps(func)
    118 def decorate_context(*args, **kwargs):
    119     with ctx_factory():
--> 120         return func(*args, **kwargs)

File ~/code/pytorch/torch/optim/lbfgs.py:462, in LBFGS.step(self, closure, zero_grad)
    457 if n_iter != max_iter:
    458     # re-evaluate function only if not in last iteration
    459     # the reason we do this: in a stochastic setting,
    460     # no use to re-evaluate that function here
    461     with torch.enable_grad():
--> 462         loss = float(closure())
    463     flat_grad = self._gather_flat_grad()
    464     opt_cond = flat_grad.abs().max() <= tolerance_grad

UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
Consider using tensor.detach() first. (Triggered internally at /home/coder/code/pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)

In [2]: import warnings
   ...: warnings.simplefilter('error')
   ...: import torch
   ...: print(torch.__version__)
   ...: a, b = torch.rand((2, 32, 32))
   ...: a.requires_grad_()
   ...: optimizer = torch.optim.LBFGS([a])
   ...: loss_fn = lambda x, y: (x-y).pow(2).mean()
   ...: 
   ...: def closure():
   ...:     optimizer.zero_grad()
   ...:     loss = loss_fn(a, b)
   ...:     loss.backward()
   ...:     return loss.item()
   ...: 
   ...: for i in range(100):
   ...:     optimizer.step(closure)
   ...:     print(i, loss_fn(a, b))
   ...: 
2.9.0a0+git23b0334
0 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
1 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
2 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
3 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
4 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
5 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
6 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
7 tensor(5.6172e-11, grad_fn=<MeanBackward0>)
...