Fixes https://github.com/pytorch/pytorch/issues/121353
our handle for `.data` in dynamo today basically just converts `y = x.data` into `y = x.detach()`. The semantics of these two ops are not quite the same, because:
(1) any future mutations on `x.data` will be fully ignored by autograd
(2) any mutations on `x.detach()` will bump x's version counter
the linked model does a .data mutation that is hidden from autograd in eager, but ends up erroring during AOTDispatcher tracing.
I updated dynamo's handling so that:
(1) when dynamo sees a call to `getattr(tensor, "data")` and calls `.detach()` we set a flag on the returned `TensorVariable` indicating it came from `.data`
(2) on any tensor method that we call with an input `TensorVariable` with this flag turned on, we proxy autograd's `preserve_version_counter` logic into the graph, to properly reset the VC after the op is run.
One thing to note is that I don't actually do this on every op that we pass the tensor to: I only do it for tensor methods that appear to be mutations (by checking for a trailing underscore). My thought was that:
(1) I didn't want to do this for **every** op that you pass `y` into, since that will e.g. triple the number of nodes in the graph, and could cause compile time regressions if you use .data
(2) this situation is pretty rare in general, and I'm hoping that "tensor method mutations" cover most reasonable mutation cases. If we manage to miss a case, you will get a loud error during tracing anyway, so there is not a safety issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131403
Approved by: https://github.com/anijain2305, https://github.com/zou3519
**Dynamo**
We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine.
The safe recipe is:
1) Disable grad
2) Call set_()
3) Manually lower the version counter on the object to hide it from the autograd engine
This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor.
**aot_autograd**
For aot_autograd, there's another snag.
Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input.
The new plan here is to:
1) Record tensor fakification policy in dynamo
2) provide a fresh fake mode to all backends
3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113080
Approved by: https://github.com/bdhirsh
