Add multi-gpu inference option

[natolambert]

Currently run_rm.py only uses one RM because RMs are not well supported generally for inference.
Current implementation is a separate run_rm_mpgu.py script. We can delete this and improve the base script if more use cases emerge.

Closes #95

[natolambert]

Inference works distributed, but I couldn't get the results gather working correctly with things like

    state.wait_for_everyone()
    
    # flatten results list of lists if is list of lists
    if state.is_main_process:
        logger.info("Gathering results")
        results = gather_object(results) # gather() is for tensors

[chrisliu298]

I have a distributed inference script for my own use case here, which might help.

The code assumes ds is a Hugging Face dataset of the following structure:

Dataset({
    features: ['chosen', 'rejected'],
    num_rows: 1000
})

and each of chosen and rejected is a list of messages in the standard chat format.

The core part of the code is shown below:

accelerator.wait_for_everyone()
with accelerator.split_between_processes(ds) as ds_shard:
    rewards = []
    for sample in tqdm(ds_shard):
        chosen_score, rejected_score = rm.get_score(sample["chosen"], sample["rejected"])
        rewards.append({"chosen_reward": chosen_score, "rejected_reward": rejected_score})
rewards_gathered = accelerate.utils.gather_object([{"rewards": rewards}])

if accelerator.is_main_process:
    all_rewards = [row for result in rewards_gathered for row in result["rewards"]]

The key steps are 1) iterate through the assigned ds_shard for the current GPU, 2) gather rewards outside the context (but for all processes), and 3) collect all rewards only in the main process.

The final all_rewards will be a list of dicts, like {"chosen_reward": 18.35, "rejected_reward": -7.89}, each corresponding to the same order of the samples in ds.