Skip to content

Distributed Training Fails #412

Description

@shirinebadi

Hi. I'm trying to run DLRM with 2 nodes each has one GPU. I use NCCL and a simple test works fine. However with DLRM I get the following issue:

[2026-03-09 23:44:20] node30:11193:11203 [0] transport/net_ib.cc:2453 NCCL WARN NET/IB: Got completion from peer 128.138.232.77<32911> with status=12 opcode=0 len=0 vendor err 129 (Send) localGid ::ffff:192.168.1.30 remoteGids::ffff:192.168.1.32 hca mlx5_0
node30:11193:11203 [0] NCCL INFO transport/net.cc:1227 -> 6

[2026-03-09 23:44:20] node30:11193:11203 [0] transport/net_ib.cc:2453 NCCL WARN NET/IB: Got completion from peer 128.138.232.77<39137> with status=12 opcode=0 len=0 vendor err 129 (Send) localGid ::ffff:192.168.1.30 remoteGids::ffff:192.168.1.32 hca mlx5_0
node30:11193:11203 [0] NCCL INFO transport/net.cc:1227 -> 6

I use A2 GPUs over ROCe. My command:

torchrun --nproc_per_node=1     --nnodes=2     --node_rank=0     --master_addr="192.168.1.30"     --master_port=1234     dlrm_s_pytorch.py     --arch-embedding-size="80000-80000-80000-80000-80000-80000-80000-80000"     --arch-sparse-feature-size=128     --arch-mlp-bot="128-128-128-128"     --arch-mlp-top="512-512-512-256-1"     --max-ind-range=40000000     --data-generation=random     --loss-function=bce     --round-targets=True     --learning-rate=1.0     --mini-batch-size=2048     --print-freq=2     --print-time     --test-freq=2     --test-mini-batch-size=2048     --memory-map     --use-gpu     --num-batches=100     --dist-backend=nccl

and node_rank=1 for the worker.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions