Tony Salomone

Add 4-bit MLX models of small models

2

Maybe some of TinyLlama, Phi, Qwen2 small models?

For some reason the Llama 3 and Llama 3.1 models can't load in to GPU before some timeout kicks in (2 minutes). I can see that there are 4 shards...

Dataset download button not not set to "Downloaded" after download

4

If you download a dataset it flips to "Downloading" when you are downloading but then when it is done it flips to "Download". Then at a later moment the screen...

Import screen doesn't show loading spinner while loading

1

The import screen shows old info while it is loading an API call after the first time. This is problematic if anything has changed!

I am hosting my server in AWS on a machine that is reporting in UTC and working on my local client in eastern time. If I look at a training...

Specifically MLX only supports some weight file formats (safetensors and nfz I think?). We currently only check architecture which means you sometimes get a "No safetensors for..." error when trying...

It looks like it is looking for local_model = true, which gets added to info.json on trianing but not on export nor on import. If we use stored_in_filesystem instead it...

Local Connection installer: No error reported if there are exceptions during "Install Dependencies" step

1

If flash_attn fails to build using the UI installer, the app will just report Success and continue. Then you'll just hit weird errors in the app later on. It would...

Generate config.json files for exported models

1

We need information in this file for RAG. Look to see if mlx-lm conversion does this now?

Control where models and other large data is stored

2

1. You can set HF_HOME, or just HF_HUB_CACHE and TransformerLab will respect this. Should you be able to change this through TransformerLab? You can also set where specific models download....