Awni Hannun
Awni Hannun
> Basically, at the top of my wish list is for this repo (or the examples repo) to contain c++-based demonstrations of running llama2 models and mistral MoE models. @dougdew64...
This looks super cool! One thing I'm wondering is what is the expected behavior when the `mx.array` goes out of scope but I still have a pointer to the memory...
This looks great to me. I don't see any obvious reasons not to go this path. It seems strictly more useful than `__array__` and keeps things pretty simple which is...
@dastrobu I think you should PR this!
``` def g(x): x_view = np.array(x, copy=False) x_view[:] += 1 # modify memory without telling mx return mx.array(x @ x) ``` That code is a pretty explicit situation (`copy=False` +...
> This was actually what started me to look into this. If buffers are exposed read only, mx.data should brute force write to them anyway. I thought about this a...
@dastrobu we discussed a bit offline and the conclusion was that we can keep buffers writable (as you have it). We'll see how it goes, if there are some hiccups...
Let's close this and open a new issue more specific to optimizing convolutional network training which is one of the major slow points still in MLX.
Great! If you have any questions about it, we can use this issue for discussion or better yet start a draft PR and use that!
Good point! Fixed in #335