gstykalin/test-dags
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Re-enable InfiniBand for distributed training

Browse Source 
Now that the dynamic master discovery is working, re-enable IB
with proper configuration for better throughput:

- NCCL_IB_DISABLE=0 (enable IB)
- Regex patterns for IB interfaces (vary by node: ibp6s0, ibp7s0)
- RoCE v2 configuration (GID_INDEX=3)
- GPUDirect RDMA enabled (NET_GDR_LEVEL=5)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
George Stykalin
2026-04-15 12:24:37 +03:00
parent 48abe40c15
commit 1c4c663431
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
dags/test-train-pytorch.py
View File

@@ -113,14 +113,18 @@ def run_training_node_func(rank, world_size):
os.environ['WORLD_SIZE'] = str(world_size) os.environ['WORLD_SIZE'] = str(world_size)
os.environ['RANK'] = str(rank) os.environ['RANK'] = str(rank)
# NCCL Configuration - Start with basic ethernet config (IB is causing issues) # NCCL Configuration for InfiniBand
os.environ['NCCL_DEBUG'] = 'INFO' os.environ['NCCL_DEBUG'] = 'INFO'
os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT) os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT)
os.environ['NCCL_BLOCKING_WAIT'] = '1' os.environ['NCCL_BLOCKING_WAIT'] = '1'
os.environ['NCCL_IB_DISABLE'] = '1' # Disable IB for now - use ethernet os.environ['NCCL_IB_DISABLE'] = '0' # Enable IB
os.environ['NCCL_SOCKET_IFNAME'] = 'eth0' # Use regular ethernet os.environ['NCCL_IB_HCA'] = '^mlx5_[0-9]+$' # Regex for Mellanox HCAs
os.environ['NCCL_SOCKET_IFNAME'] = '^ibp[0-9]+s[0-9]+$' # Regex for IB interfaces
os.environ['NCCL_P2P_DISABLE'] = '0' # Enable P2P os.environ['NCCL_P2P_DISABLE'] = '0' # Enable P2P
os.environ['NCCL_IGNORE_CPU_AFFINITY'] = '1' # Better compatibility os.environ['NCCL_IGNORE_CPU_AFFINITY'] = '1' # Better compatibility
os.environ['NCCL_NET_GDR_LEVEL'] = '5' # Enable GPUDirect RDMA
os.environ['NCCL_IB_GID_INDEX'] = '3' # Use RoCE v2 (common for modern IB)
os.environ['NCCL_IB_TC'] = '160' # Traffic class for low latency
print(f"[{rank}] Environment configured:") print(f"[{rank}] Environment configured:")
print(f" MASTER_ADDR: {MASTER_ADDR}") print(f" MASTER_ADDR: {MASTER_ADDR}")