Amog Kamsetty

Ah yes, that's right- it can also be UUIDs. > GPU identifiers are given as integer indices or as UUID strings. From https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars. In this case, then we should always...

@pcmoritz thoughts about the proposed change to always return strings?

@dumpmemory what are the advantages of deepspeed vs. fsdp?

Is this ready to be merged in?

Perhaps try out compress gpt: https://github.com/yasyf/compress-gpt It seems to be built for exactly this use case.

Hey @reciprocated, @LouisCastricato, @ayulockin-- circling back on this thread. As @ayulockin mentions, you can use `tune.with_resources` to allocate multiple GPUs per trial, but the challenge is that you need a...

Some torchvision transforms work on a batched input. Should we allow users to also specify a separate batched_transform for better performance? See our batch prediction benchmarks for an example

API could be something like this: ``` single_transform = transforms.ToTensor() batch_transform = transforms.Compose([CenterCrop(...), Normalize(...)]) preprocessor = TorchVisionPreprocessor(transform=single_transform, batch_transform=batch_transform) ``` Single preprocessor that can accept both options. `batch_transform` arg is Optional.

Are we still planning to remove `iter_batches` from the base Dataset API and instead have `ds.to_iterator().iter_batches()`? Basically avoiding duplicate APIs in both `Dataset` and `DatasetIterator`

This was an intentional change previously as `ray.wait()` time is not interpretable to the user. From the user perspective, how are they supposed to interpret this?