Share model

[angeloskath]

Mostly putting this to check if people think it's useful. We could also put it in mlx_lm.manage.

This PR introduces a utility called mlx_lm.share that can be used to broadcast models (or any folder/file) from 1 node to the others using mlx.core.distributed. It is often the case that we download a model on one node and then have to move it to others via external storage or something similar.

Thunderbolt 5 with JACCL actually saturates the write bandwidth of the disk so with mlx_lm.share I am getting 5-6 GB/s broadcast so no need for external storage to move things around. It would be strictly slower...

Examples

# Assumes we have ran
# mlx.distributed_config --hosts h1,h2,h3,h4,... --over thunderbolt \
#                        --backend jaccl (or jaccl-ring) --output hosts.json

mlx_lm.share --path mlx-community/Kimi-K2.5 --hostfile hosts.json

mlx_lm.share --path my-big-file.bin --hostfile hosts.json --dst /tmp/dst.bin

[awni]

nit: wdyt about renaming that to --model to be consistent with other command?

[awni]

Although if the intention is this can be used for anything (not just a model) than --path is probably better.

[awni]

Very nice! This seems quite useful!

[ivanfioravanti]

Nice! No more manual copy and paste across nodes!