[Train] Using Ray Train with fractional GPUs leads to NCCL error - Duplicate GPU detected

[choosehappy]

What happened + What you expected to happen

I have a machine with 2 GPUs. In my use case, I need to run multiple TorchTrainers concurrently on different datasets.

Ray Train is setup to avoid fragmenting jobs across different GPUs, but in this case this leads to an obvious NCCL error:

Duplicate GPU detected : rank 0 and rank 1 both on CUDA device 3b000

I would expect that when using Ray Train, the TorchTrainer internally knows that it should provide CUDA_VISIBLE_DEVICES equivalent to a "spread" across GPUs as opposed to a "pack" onto a single GPU.

I can get "something" that looks like the correct behavior if launching the first train job with GPU:.6, so it spreads across both jobs, and then launching a second train job with GPU:.4.

However, if you run a job with anything < .5, then both workers will try to use the same GPU resulting in the NCCL error.

I tried playing aggressively with something like:

cuda_dev=torch.device('cuda',train.get_context().get_local_rank())
model=ray.train.torch.prepare_model(model,cuda_dev)

but internally the Ray device_manager really hates this and there doesn't seem to be working combination that allows it to slide through.

Any thoughts?

Versions / Dependencies

Ray 2.37, Python 3.10

Reproduction script

import ray
from ray.train import ScalingConfig
from torchvision.models import resnet18
import ray.train.torch

ray.init()


def trainpred_func(config):
    model = resnet18(num_classes=10)
    model=ray.train.torch.prepare_model(model)
    

scaling_config = ray.train.ScalingConfig(num_workers=2, use_gpu=True, resources_per_worker={"GPU":.1})
trainer = ray.train.torch.TorchTrainer(trainpred_func,scaling_config=scaling_config)
trainer.fit()

Issue Severity

High: It blocks me from completing my task.

[dhirajtobii]

Does Ray support fractional scaling config for gpu workers with default nccl backend. I can see that above code works fine with gloo backend

I am also interested in knowing the status of this feature

[choosehappy]

Coming back to this, i found a janky work around:

1. create 1 docker instance per GPU on a node

start within the first container:

CUDA_VISIBLE_DEVICES=0 ray start --address='172.20.0.2:6379'

and then within the second container:

CUDA_VISIBLE_DEVICES=1 ray start --address='172.20.0.2:6379'

etc etc

2. this essentially creates multiple "nodes", each with 1 GPU
3. and thus one can request a placement_strategy of SPREAD, which will avoid the packing behavior described above:

scaling_config = ray.train.ScalingConfig(num_workers=2, use_gpu=True, resources_per_worker={"GPU":.1}, placement_strategy="SPREAD")

obviously this is a terrible idea for multiple reasons, for example, the total CPU count and RAM count, is 2x the correct amount. one can hack around that as well. one can also imagine a setting where the scheduler cannot SPREAD, and leave the system back in a fail state. STRICT_SPREAD is likely a smarter option to ensure failure if the spread wouldn't take place

but at least it "works" : )

[sud474]

Hi, Did you find a solution to this?

[choosehappy]

only the janky one i mentioned above