[dynamo,torch_function] __torch_function__ does not respect kwargs

[lezcano]

🐛 Describe the bug

Found when debugging some new failures coming from #117625.

Our implementation of torch_function does not trace the call within if has_torch_function in eager. This makes kwargs not to be populated correctly. For example in

pytorch/torch/nn/functional.py

Lines 2232 to 2243 in c378001

	if has_torch_function_variadic(input, weight):
	return handle_torch_function(
	embedding,
	(input, weight),
	input,
	weight,
	padding_idx=padding_idx,
	max_norm=max_norm,
	norm_type=norm_type,
	scale_grad_by_freq=scale_grad_by_freq,
	sparse=sparse,
	)

we see that some kwargs are populated, even when all the args are passed as positional args. This a few calls further down where the implementation expects just 2 positional args:

pytorch/torch/nn/utils/_expanded_weights/embedding_expanded_weights.py

Line 16 in c378001

ctx.input, ctx.weight = expanded_args

cc @hameerabbasi @rgommers @ezyang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @peterbell10 @aakhundov @mlazos

Versions

master