pytorch - 💡(How to fix) Fix torch.nn.functional.scaled_grouped_mm doesn't support compile

E               torch._dynamo.exc.Unsupported: Unexpected type in sourceless builder
E                 Higher Order Operator: autograd.Function
E                 Explanation: SourcelessBuilder.create does not know how to wrap <class 'torch.nn.functional._ScalingType'>
E                 Hint: This is likely to be a Dynamo bug. Please report an issue to PyTorch.
E               
E                 Developer debug context: torch.nn.functional._ScalingType
E               
E                For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0116.html
E               
E               from user code:
E                  File "/home/dev/ao/torchao/prototype/moe_training/mxfp8_grouped_mm.py", line 136, in _to_mxfp8_then_scaled_grouped_mm
E                   output = _MXFP8GroupedMM.apply(
E                 File "/home/dev/.conda/envs/torch/lib/python3.13/site-packages/torch/_functorch/autograd_function.py", line 880, in inner
E                   outs = orig_fwd(*args, **kwargs)
E                 File "/home/dev/ao/torchao/prototype/moe_training/mxfp8_grouped_mm.py", line 245, in forward
E                   output = _compute_fwd(
E                 File "/home/dev/ao/torchao/prototype/moe_training/mxfp8_grouped_mm.py", line 408, in _compute_fwd
E                   return _compute_fwd_sm100(
E                 File "/home/dev/ao/torchao/prototype/moe_training/mxfp8_grouped_mm.py", line 590, in _compute_fwd_sm100
E                   ScalingType.BlockWise1x32,

---

E       NotImplementedError: aten::_scaled_grouped_mm_v2: attempted to run this operator with Meta tensors, but there was no fake impl or Meta kernel registered. You may have run into this message while using an operator with PT2 compilation APIs (torch.compile/torch.export); in order to use this operator with those APIs you'll need to add a fake impl. Please see the following for next steps:  https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html

---

E                 Higher Order Operator: autograd.Function
E                 Explanation: Dynamo does not know how to trace the builtin `torch._VariableFunctionsClass._scaled_grouped_mm_v2.` This function is either a Python builtin (e.g. _warnings.warn) or a third-party C/C++ Python extension (perhaps created with pybind).