AIR, TorchTrainer, DDP and NCCL Timeout

Ray Libraries (Data, Train, Tune, Serve)
1

I have a 4 GPU single node and about 1.5 - 2 hours into training, I get this error:

[Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=50966, OpType=BROADCAST, Timeout(ms)=3600000) ran for 3609559 milliseconds before timing out.

mgpu1
mgpu13089×657 204 KB

I have tried increasing the timeout and still get the same error. I noticed this thread with similar issues: TorchTrainer: Collective operation timeout: WorkNCCL - #2 by saivivek15, but I do not have NVLink on my node.

What other options are available?

2

Is the job training successfully up until this point? Is it hanging up until that point? Are you able to correlate the failure point (or the failure point - 3600000ms) to some event?

3

Indeed, it is training, reporting a decreasing loss, etc but it fails generally around that point. There is quite a bit of discussion about these types of errors in the Torch DDP community, but I am wondering if there tends to be an obvious thing to try or not.