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

Revert to ethernet - IB PKEY misconfigured in subnet manager

Browse Source 
Root cause: Subnet manager assigns different PKEYs to ports
- test-1 (LID 0x44): Full membership PKEYs (0xa001, 0x9001)
- test-2 (LID 0x7d): Limited membership PKEY 0x7fff only

This causes MTU mismatch (4092 vs 2044) and NCCL failure.
Cluster admin needs to configure opensm partitions to fix.

Ethernet is working and stable.

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
George Stykalin
2026-04-15 18:45:54 +03:00
parent c659cd2f4b
commit 6f88661b8b
1 changed files with 6 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
dags/test-train-pytorch.py
View File

@@ -40,17 +40,13 @@ def run_training_node_func(rank, world_size):
import time import time
# CRITICAL: Set NCCL environment variables BEFORE importing torch # CRITICAL: Set NCCL environment variables BEFORE importing torch
# This ensures NCCL reads these settings during initialization # NOTE: IB is disabled due to subnet manager PKEY mismatch (test-2 only has limited membership 0x7fff)
# Cluster admin needs to configure opensm partitions to assign full membership PKEYs to both ports
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['TORCH_NCCL_BLOCKING_WAIT'] = '1' # Updated naming os.environ['TORCH_NCCL_BLOCKING_WAIT'] = '1' # Updated naming
os.environ['NCCL_IB_DISABLE'] = '0' # Enable IB os.environ['NCCL_IB_DISABLE'] = '1' # Disable IB - use ethernet (PKEY misconfiguration)
os.environ['NCCL_SOCKET_IFNAME'] = 'eth0' # Fallback to eth0 os.environ['NCCL_SOCKET_IFNAME'] = 'eth0' # Use ethernet (working)
os.environ['NCCL_IB_GID_INDEX'] = '3' # Use RoCE v2 (GID index 3)
os.environ['NCCL_IB_TC'] = '106' # Traffic class for RoCE
os.environ['NCCL_IB_TIMEOUT'] = '22' # IB timeout (higher = more tolerant)
os.environ['NCCL_IB_RETRY_CNT'] = '7' # IB retry count
os.environ['NCCL_NET_GDR_LEVEL'] = '5' # Enable GPUDirect RDMA
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
@@ -137,9 +133,8 @@ def run_training_node_func(rank, world_size):
print(f" MASTER_PORT: {MASTER_PORT}") print(f" MASTER_PORT: {MASTER_PORT}")
print(f" RANK: {rank}") print(f" RANK: {rank}")
print(f" WORLD_SIZE: {world_size}") print(f" WORLD_SIZE: {world_size}")
print(f" NCCL_IB_DISABLE: {os.environ.get('NCCL_IB_DISABLE')} (IB enabled)") print(f" NCCL_IB_DISABLE: {os.environ.get('NCCL_IB_DISABLE')} (ethernet mode)")
print(f" NCCL_IB_GID_INDEX: {os.environ.get('NCCL_IB_GID_INDEX')}") print(f" NCCL_SOCKET_IFNAME: {os.environ.get('NCCL_SOCKET_IFNAME')}")
print(f" NCCL_IB_TIMEOUT: {os.environ.get('NCCL_IB_TIMEOUT')}")
# STEP 3.5: Pre-flight checks # STEP 3.5: Pre-flight checks
print(f"[{rank}] Running pre-flight checks...") print(f"[{rank}] Running pre-flight checks...")