search

ggerganov / ggml

Tensor library for machine learning
MIT License
10.88k stars 1k forks source link

`ggml_flip` or `ggml_pad_reflect`? #819

Open PABannier opened 4 months ago

PABannier commented 4 months ago

Hello,

I have implemented a custom operation ggml_pad_reflec_1d on my ggml fork. This is required for Encodec.cpp. Should I upstream this operation? I would write a ggml_pad_reflect supporting 1D and 2D input, as in PyTorch's nn.ReflectionPad2d

Alternatively, I was considering implementing a ggml_flip operation (PyTorch's flip). This would allow us to implement a pad_reflect operation directly in 'user space' using flipping instead of adding a quite niche operation (pad_reflect) to ggml.

What are your thoughts? @slaren @ggerganov

ggerganov commented 4 months ago

Alternatively, I was considering implementing a ggml_flip operation (PyTorch's flip). This would allow us to implement a pad_reflect operation directly in 'user space' using flipping instead of adding a quite niche operation (pad_reflect) to ggml.

I don't have a good idea which option would be more useful, so whatever you think makes more sense - PR welcome

balisujohn commented 3 months ago

I like ggml_pad_reflect as I need this functionality for tortoise.cpp. @PABannier can I have your permission to turn that code you have there into a PR for ggml so I can get this quickly upstreamed(I don't want to take your credit for the PR without asking).

balisujohn commented 3 months ago

I ask because I need a working version of this to proceed with tortoise so I'm going to have to integrate it into at least my local copy of ggml anyway.

PABannier commented 3 months ago

@balisujohn Please go on! :)