ml-explore
Possible to reduce disc writes when converting models?
Hello!
I am regularly converting models to MLX and uploading them to HF. When I find a model I upload both q3, q4, q6, q8 and full mlx version.
This is causing a lot of disc writes and I'm kind of worried that I'm speedrunning my new macbook that has 500gb ssd to failure (I've written and deleted several tb of models in last 2 weeks).
Would it be possible to somehow not save the output of the conversion to disc and just kind of convert it and upload straight to HF? Or somehow run this on an external drive or something? Or something else that would minimize how much data is written to the disc when converting multiple quants of the same model?
I'm not sure exactly of the details of how it is converting and saving it to disc but would much appreciate some help with this as its a fun hobby for me to convert models.
Thanks for making so many MLX models, that's super awesome!
Or somehow run this on an external drive or something?
That should already work if you specify the correct path.
You could also consider using the conversion space on Hugging Face. You don't even need to download anything for that to work :).
Wow thanks, did not know about that conversion space. Just tried it and it worked :D That means I can convert bigger models than my 16gb ram allows too... This is huge, they are not ready for me 🤠
Oh but its not possible to upload a non quantized version it seems like? Cause when I upload the complete MLX version non-quantized that usually gets as many downloads as the quantized ones
Oh but its not possible to upload a non quantized version it seems like? Cause when I upload the complete MLX version non-quantized that usually gets as many downloads as the quantized ones
PR in progress https://huggingface.co/spaces/mlx-community/mlx-my-repo/discussions/41
