Summary: 57 instances, 54 unique

Text	Count
// TODO: This is kinda complicated. Saving this for a future date.	1
// TODO: linear is pretty important for performance, but I'm not sure how to work	1
# # TODO: make sure PyTorch has tanh_backward implemented for jvp!!	1
throw std::runtime_error("TODO: implement record function");	1
# TODO: remove the monkeypatching and add an extension point into PyTorch core	1
// TODO: this is technically kind of sketchy, since we're relying on the fact	1
// TODO: need to reset sizes/strides on mutation	1
// TODO: Some duplication with torch/csrc/autograd/python_variable.cpp	1
// TODO: we only need to do the following (marked with !) on in-place functions	1
// TODO: replace with targetable functionalization	1
), "TODO: support backwards for broadcasting"	1
# TODO: Remove the following hack for namedtuples	1
// TODO: upstream into core	1
// TODO: is this lock safe? No one else should be writing to the same bucket	1
// TODO: https://github.com/pytorch/pytorch/issues/69991	1
// TODO: figure out what this is	1
assert spec.out == 0, "TODO: support autograd on out= ?"	1
// TODO: First entry in the stack is a default autograd key.	2
# TODO: forward-mode AD: does it also respect no_grad? What does that mean	1
// TODO: Still a problem with composabiilty and AutoNonVariableTypeGuard.	1
//   // TODO: No clue if this works if the first non-batch dim isn't size 1	1
// TODO: I don't think this ctor gets used.	1
// TODO: if the result isn't batched but bias is, then we need to do the following.	1
# TODO: it looks like the autograd alias key doesn't work	1
// NB: not lock safe. TODO: does it need a lock?	1
assert spec.alias_group == 0, "TODO: support aliased backwards"	1
// TODO: actually populate kwargs sometimes?  At the moment, every argument	1
throw std::runtime_error("TODO: support overloaded signatures");	1
// TODO: might be more efficient if we scan first then not copy? Depends.	1
throw std::runtime_error("TODO: support other arg counts");	1
// TODO: all of these should be fixed in a more blessed way. In particular,	1
// TODO: Get rid of this, it is bad for composability	1
// TODO: can replace with better conditional functionalization	1
// TODO: generalize this to more transforms	1
// TODO: is this right?	1
assert layout == torch.strided, "TODO: support other layouts"	1
# TODO: The following shows that the per-sample-grads computed are different.	1
# TODO - move this to PyTorch core. This overrides the pytree implementation for	1
// TODO: is this even right?	2
# TODO: some namedtuple return. Also, testing for codegen	1
throw std::runtime_error("TODO: handle more dims");	1
# TODO: We don't need to copy the model to create a stateless copy	2
// TODO: delete the following after confirming performance	1
// TODO: The following algorithm only works for batch dim == 0.	1
assert s.alias_group >= 0, "TODO: support complex aliasing"	1
// TODO: A little bird says that unfold + matmul is actually faster than	1
# TODO - move the hashing part of static_args to C++.	1
// TODO: this is a weird special case...	1
// TODO: Probably better way to metaprogram this	1
// TODO: Once the divergent behavior for 0-d scalar is fixed, we should use REDUCTION_BOXED_ARGS	1
throw std::runtime_error("TODO: support more than one output");	1
), "TODO: support aliased backwards"	1
// TODO: Consider rewriting the following to look like:	1
// TODO: dedup	1