Awni Hannun
Awni Hannun
@jdeschena do you intend to come back to this PR? I think it could be nice to have basic RNN support but if so would be good to move this...
Looks like a format failed? Would you mind checking it?
Wow that would be amazing! Are you comfortable writing Metal / kernels on the GPU? Just to be sure, it's a pretty big task, if you haven't done GPU programming,...
It is indeed a commonly used method. Just thinking out loud: what do you think about `Module.astype`? Also torch has [`Module.to`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module.to) which is I think is the more commonly used...
@altaic this is at the `nn.Module` level. We already have `astype` at the array level :). I find the consistency with that pleasing, but I'm curious on other's thoughts.
That is a very good point about in-place vs copy.
In my opinion `set_dtype` > `setdtype` > `apply_dtype` (which I don't really like).
Unless anyone feels differently, I would suggests we go with `set_dtype`. You've convinced me it's the best option. Would be great if you would send a PR. And thank you...
You need to be careful with timing because stuff is asynchronous. So I would use an `mx.eval` right before each time measurement to make sure everything is done executing. Once...
Will take a look at the DQN bit. But Just FYI in you other test, it makes sense that if you remove the call to `.item` things will seem faster....