George Stykalin
ebb3284853
The regex patterns for IB interfaces weren't working. Now letting NCCL auto-detect the best available transport (IB or ethernet). - Removed NCCL_IB_DISABLE (let NCCL decide) - Removed NCCL_SOCKET_IFNAME (let NCCL auto-detect) - Kept RoCE v2 configuration Co-Authored-By: Claude <noreply@anthropic.com>
347 lines
11 KiB
Python
347 lines
11 KiB
Python
from airflow import DAG
|
|
from airflow.operators.python import PythonOperator
|
|
from airflow.models import Variable
|
|
import pendulum
|
|
from datetime import timedelta
|
|
import json
|
|
|
|
# --- CONFIGURATION ---
|
|
WORLD_SIZE = 2
|
|
MASTER_ADDR = "airflow-worker-gpu-0.airflow-worker-gpu.george-tests.svc.cluster.local"
|
|
MASTER_PORT = "29500"
|
|
NCCL_TIMEOUT = 1800
|
|
NAMESPACE = "george-tests"
|
|
|
|
default_args = {
|
|
'owner': 'airflow',
|
|
'retries': 1,
|
|
'retry_delay': timedelta(minutes=2),
|
|
'execution_timeout': timedelta(hours=2),
|
|
}
|
|
|
|
|
|
def prepare_training_func():
|
|
"""Initialize shared state for synchronization"""
|
|
sync_state = {
|
|
'ready_workers': [],
|
|
'training_started': False,
|
|
'start_time': None
|
|
}
|
|
Variable.set('ddp_sync_state', json.dumps(sync_state))
|
|
print("Training preparation complete. Sync state initialized.")
|
|
return True
|
|
|
|
|
|
def run_training_node_func(rank, world_size):
|
|
"""Execute distributed training with synchronization"""
|
|
import os
|
|
import socket
|
|
import torch
|
|
import torch.distributed as dist
|
|
import torch.nn as nn
|
|
import torch.optim as optim
|
|
from datetime import datetime
|
|
import time
|
|
|
|
print(f"{'='*60}")
|
|
print(f"Node Rank {rank}/{world_size} - Starting at {datetime.now()}")
|
|
print(f"Hostname: {socket.gethostname()}")
|
|
print(f"{'='*60}")
|
|
|
|
# STEP 1: Signal that this worker is ready
|
|
print(f"[{rank}] Signaling ready state...")
|
|
max_wait = 300 # 5 minutes
|
|
start_wait = time.time()
|
|
|
|
while time.time() - start_wait < max_wait:
|
|
try:
|
|
sync_state = json.loads(Variable.get('ddp_sync_state', default_var='{}'))
|
|
|
|
if rank not in sync_state.get('ready_workers', []):
|
|
sync_state.setdefault('ready_workers', []).append(rank)
|
|
Variable.set('ddp_sync_state', json.dumps(sync_state))
|
|
print(f"[{rank}] Marked as ready. Ready workers: {sync_state['ready_workers']}")
|
|
|
|
# STEP 2: Wait for all workers to be ready
|
|
if len(sync_state.get('ready_workers', [])) == world_size:
|
|
print(f"[{rank}] All {world_size} workers are ready! Proceeding to training...")
|
|
break
|
|
|
|
print(f"[{rank}] Waiting for other workers... ({len(sync_state.get('ready_workers', []))}/{world_size} ready)")
|
|
time.sleep(2)
|
|
|
|
except Exception as e:
|
|
print(f"[{rank}] Error during sync: {e}")
|
|
time.sleep(2)
|
|
else:
|
|
raise RuntimeError(f"[{rank}] Timeout waiting for all workers to be ready!")
|
|
|
|
# STEP 2.5: Rank 0 stores its address, others wait for it
|
|
print(f"[{rank}] Determining master address...")
|
|
my_fqdn = socket.getfqdn()
|
|
print(f"[{rank}] My FQDN: {my_fqdn}")
|
|
|
|
if rank == 0:
|
|
# Rank 0 stores its address as the master
|
|
sync_state = json.loads(Variable.get('ddp_sync_state', default_var='{}'))
|
|
sync_state['master_addr'] = my_fqdn
|
|
Variable.set('ddp_sync_state', json.dumps(sync_state))
|
|
print(f"[{rank}] I am the master. Stored my address: {my_fqdn}")
|
|
else:
|
|
# Other ranks wait for rank 0 to store its address
|
|
max_wait = 60
|
|
start_wait = time.time()
|
|
while time.time() - start_wait < max_wait:
|
|
sync_state = json.loads(Variable.get('ddp_sync_state', default_var='{}'))
|
|
if 'master_addr' in sync_state:
|
|
master_addr = sync_state['master_addr']
|
|
print(f"[{rank}] Master address retrieved: {master_addr}")
|
|
break
|
|
print(f"[{rank}] Waiting for master address...")
|
|
time.sleep(1)
|
|
else:
|
|
raise RuntimeError(f"[{rank}] Timeout waiting for master address!")
|
|
|
|
# Small delay to ensure all workers see the ready state
|
|
time.sleep(3)
|
|
|
|
# STEP 3: Configure distributed environment
|
|
# Use dynamic master address from rank 0
|
|
actual_master_addr = my_fqdn if rank == 0 else sync_state.get('master_addr', MASTER_ADDR)
|
|
os.environ['MASTER_ADDR'] = actual_master_addr
|
|
os.environ['MASTER_PORT'] = MASTER_PORT
|
|
os.environ['WORLD_SIZE'] = str(world_size)
|
|
os.environ['RANK'] = str(rank)
|
|
|
|
# NCCL Configuration - Let NCCL auto-detect best transport
|
|
os.environ['NCCL_DEBUG'] = 'INFO'
|
|
os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT)
|
|
os.environ['NCCL_BLOCKING_WAIT'] = '1'
|
|
# Don't disable IB - let NCCL auto-detect
|
|
# Don't set SOCKET_IFNAME - let NCCL find the best interface
|
|
os.environ['NCCL_P2P_DISABLE'] = '0' # Enable P2P
|
|
os.environ['NCCL_IGNORE_CPU_AFFINITY'] = '1' # Better compatibility
|
|
os.environ['NCCL_IB_GID_INDEX'] = '3' # Use RoCE v2
|
|
os.environ['NCCL_IB_TC'] = '106' # Traffic class (lower value for compatibility)
|
|
os.environ['NCCL_IB_TIMEOUT'] = '22' # IB timeout in microseconds
|
|
|
|
print(f"[{rank}] Environment configured:")
|
|
print(f" MASTER_ADDR: {MASTER_ADDR}")
|
|
print(f" MASTER_PORT: {MASTER_PORT}")
|
|
print(f" RANK: {rank}")
|
|
print(f" WORLD_SIZE: {world_size}")
|
|
|
|
# STEP 3.5: Pre-flight checks
|
|
print(f"[{rank}] Running pre-flight checks...")
|
|
print(f"[{rank}] Hostname: {socket.gethostname()}")
|
|
print(f"[{rank}] FQDN: {socket.getfqdn()}")
|
|
|
|
# Test DNS resolution
|
|
try:
|
|
import subprocess
|
|
result = subprocess.run(['getent', 'hosts', MASTER_ADDR], capture_output=True, text=True)
|
|
print(f"[{rank}] DNS resolution for {MASTER_ADDR}:")
|
|
print(f" {result.stdout.strip()}")
|
|
except Exception as e:
|
|
print(f"[{rank}] Warning: Could not verify DNS: {e}")
|
|
|
|
# Check network interfaces
|
|
try:
|
|
result = subprocess.run(['ip', 'addr'], capture_output=True, text=True)
|
|
ib_lines = [line for line in result.stdout.split('\n') if 'ibp' in line or 'ib0' in line]
|
|
if ib_lines:
|
|
print(f"[{rank}] Found IB interfaces:")
|
|
for line in ib_lines[:5]:
|
|
print(f" {line.strip()}")
|
|
except Exception as e:
|
|
print(f"[{rank}] Warning: Could not list interfaces: {e}")
|
|
|
|
# STEP 4: Initialize process group
|
|
print(f"[{rank}] Initializing process group (backend=nccl)...")
|
|
print(f"[{rank}] This may take 30-60 seconds...")
|
|
print(f"[{rank}] NCCL_DEBUG={os.environ.get('NCCL_DEBUG', 'not set')}")
|
|
print(f"[{rank}] NCCL_IB_DISABLE={os.environ.get('NCCL_IB_DISABLE', 'not set')}")
|
|
print(f"[{rank}] NCCL_SOCKET_IFNAME={os.environ.get('NCCL_SOCKET_IFNAME', 'not set')}")
|
|
|
|
# Flush output to ensure logs are visible
|
|
import sys
|
|
sys.stdout.flush()
|
|
|
|
try:
|
|
dist.init_process_group(
|
|
backend="nccl",
|
|
init_method="env://",
|
|
timeout=timedelta(minutes=10),
|
|
rank=rank,
|
|
world_size=world_size
|
|
)
|
|
print(f"[{rank}] ✓ Successfully joined distributed group!")
|
|
print(f" Process group size: {dist.get_world_size()}")
|
|
print(f" My rank: {dist.get_rank()}")
|
|
except Exception as e:
|
|
print(f"[{rank}] ✗ Failed to initialize process group!")
|
|
print(f" Error type: {type(e).__name__}")
|
|
print(f" Error: {str(e)}")
|
|
import traceback
|
|
traceback.print_exc()
|
|
raise
|
|
|
|
# STEP 5: Setup GPU
|
|
device = torch.device("cuda:0")
|
|
torch.cuda.set_device(device)
|
|
|
|
print(f"[{rank}] GPU Configuration:")
|
|
print(f" Device: {torch.cuda.get_device_name(0)}")
|
|
print(f" Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.2f} GB")
|
|
|
|
# STEP 6: Define model
|
|
model = nn.Sequential(
|
|
nn.Linear(10, 128),
|
|
nn.ReLU(),
|
|
nn.Linear(128, 128),
|
|
nn.ReLU(),
|
|
nn.Linear(128, 10)
|
|
).to(device)
|
|
|
|
ddp_model = nn.parallel.DistributedDataParallel(
|
|
model,
|
|
device_ids=[0],
|
|
output_device=0
|
|
)
|
|
|
|
criterion = nn.MSELoss()
|
|
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001, momentum=0.9)
|
|
|
|
print(f"[{rank}] Model initialized with {sum(p.numel() for p in model.parameters())} parameters")
|
|
|
|
# STEP 7: Training loop
|
|
print(f"\n[{rank}] {'='*60}")
|
|
print(f"[{rank}] Starting Training Loop")
|
|
print(f"[{rank}] {'='*60}")
|
|
|
|
num_epochs = 10
|
|
batch_size = 32
|
|
num_batches = 5
|
|
|
|
for epoch in range(num_epochs):
|
|
ddp_model.train()
|
|
epoch_loss = 0.0
|
|
|
|
for batch_idx in range(num_batches):
|
|
torch.manual_seed(epoch * num_batches + batch_idx)
|
|
inputs = torch.randn(batch_size, 10).to(device)
|
|
labels = torch.randn(batch_size, 10).to(device)
|
|
|
|
optimizer.zero_grad()
|
|
outputs = ddp_model(inputs)
|
|
loss = criterion(outputs, labels)
|
|
loss.backward()
|
|
optimizer.step()
|
|
|
|
epoch_loss += loss.item()
|
|
|
|
avg_loss = epoch_loss / num_batches
|
|
|
|
# Synchronize loss across ranks
|
|
loss_tensor = torch.tensor([avg_loss]).to(device)
|
|
dist.all_reduce(loss_tensor, op=dist.ReduceOp.AVG)
|
|
global_avg_loss = loss_tensor.item()
|
|
|
|
if rank == 0:
|
|
print(f"[{rank}] Epoch {epoch+1}/{num_epochs} | Global Avg Loss: {global_avg_loss:.6f}")
|
|
|
|
print(f"\n[{rank}] {'='*60}")
|
|
print(f"[{rank}] Training Complete!")
|
|
print(f"[{rank}] {'='*60}")
|
|
|
|
# STEP 8: Cleanup
|
|
dist.destroy_process_group()
|
|
print(f"[{rank}] Process group destroyed. Finished at {datetime.now()}")
|
|
|
|
return {
|
|
'rank': rank,
|
|
'final_loss': global_avg_loss if rank == 0 else avg_loss,
|
|
'epochs_completed': num_epochs,
|
|
'status': 'success'
|
|
}
|
|
|
|
|
|
def cleanup_sync_state_func():
|
|
"""Clean up synchronization state"""
|
|
try:
|
|
Variable.delete('ddp_sync_state')
|
|
print("Sync state cleaned up")
|
|
except:
|
|
pass
|
|
return True
|
|
|
|
|
|
def training_summary_func(**context):
|
|
"""Aggregate and display training results"""
|
|
ti = context['ti']
|
|
|
|
print(f"\n{'='*60}")
|
|
print(f"DISTRIBUTED TRAINING SUMMARY")
|
|
print(f"{'='*60}")
|
|
|
|
for rank in range(WORLD_SIZE):
|
|
try:
|
|
result = ti.xcom_pull(task_ids=f"train_rank_{rank}")
|
|
if result and result.get('status') == 'success':
|
|
print(f"Rank {result['rank']}: ✓ Completed {result['epochs_completed']} epochs")
|
|
print(f" Final loss: {result['final_loss']:.6f}")
|
|
except Exception as e:
|
|
print(f"Warning: Could not get result for rank {rank}: {e}")
|
|
|
|
print(f"\n✓ Training completed!")
|
|
return {'status': 'success', 'workers': WORLD_SIZE}
|
|
|
|
|
|
with DAG(
|
|
dag_id='pytorch_distributed_training_ddp_production',
|
|
default_args=default_args,
|
|
schedule=None,
|
|
start_date=pendulum.today('UTC').add(days=-1),
|
|
catchup=False,
|
|
tags=['gpu', 'ml', 'distributed'],
|
|
max_active_runs=1,
|
|
max_active_tasks=10,
|
|
) as dag:
|
|
|
|
# Preparation task
|
|
prep = PythonOperator(
|
|
task_id='prepare_training',
|
|
python_callable=prepare_training_func,
|
|
queue='gpu'
|
|
)
|
|
|
|
# Training tasks - one per rank
|
|
training_tasks = []
|
|
for i in range(WORLD_SIZE):
|
|
task = PythonOperator(
|
|
task_id=f'train_rank_{i}',
|
|
python_callable=run_training_node_func,
|
|
op_kwargs={'rank': i, 'world_size': WORLD_SIZE},
|
|
queue='gpu',
|
|
)
|
|
training_tasks.append(task)
|
|
|
|
# Cleanup task
|
|
cleanup = PythonOperator(
|
|
task_id='cleanup_sync_state',
|
|
python_callable=cleanup_sync_state_func,
|
|
queue='gpu',
|
|
trigger_rule='all_done'
|
|
)
|
|
|
|
# Summary task
|
|
summary = PythonOperator(
|
|
task_id='training_summary',
|
|
python_callable=training_summary_func,
|
|
queue='gpu',
|
|
trigger_rule='all_done'
|
|
)
|
|
|
|
# Set dependencies
|
|
prep >> training_tasks >> cleanup >> summary
|