[feature request] support input/output to fsspec path

[leoleoasd]

Support loading model from s3 and save checkpoint to fsspec path.

[ebsmothers]

Hi @leoleoasd thanks for creating the issue. @joecummings is working on updating some of our checkpointing abstractions, so he may have some thoughts on this.

[leoleoasd]

I researched a little, and I found that safetensor's safeopen uses rust's File::open and MMAP:

https://github.com/huggingface/safetensors/blob/e61e87240d0eabc9749a67ccebe38dca620d48b4/bindings/python/src/lib.rs#L396-L399

So this may not be possible with safetensor?

[joecummings]

I'll have to dig a little deeper on safetensor specifically, but I think the PyTorch Checkpointing team has worked on enabling S3 read/write through use of DCP StorageReaders and StorageWriters: https://pytorch.org/docs/main/distributed.checkpoint.html#torch.distributed.checkpoint.StorageReader

@saumishr / @ankitageorge Is there a good resource for how to create your own adapter and what might be needed for S3?