update dags
This commit is contained in:
@@ -1,9 +1,9 @@
|
||||
from airflow import DAG
|
||||
from airflow.operators.python import PythonOperator
|
||||
from airflow.models import Variable
|
||||
import pendulum
|
||||
from datetime import timedelta
|
||||
import json
|
||||
import os
|
||||
import time
|
||||
|
||||
# --- CONFIGURATION ---
|
||||
WORLD_SIZE = 2
|
||||
@@ -11,6 +11,8 @@ MASTER_ADDR = "airflow-worker-gpu-0.airflow-worker-gpu"
|
||||
MASTER_PORT = "29500"
|
||||
NCCL_TIMEOUT = 1800
|
||||
|
||||
BARRIER_FILE = f"/tmp/ddp_barrier_{WORLD_SIZE}"
|
||||
|
||||
default_args = {
|
||||
'owner': 'airflow',
|
||||
'retries': 1,
|
||||
@@ -20,83 +22,78 @@ default_args = {
|
||||
|
||||
|
||||
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.")
|
||||
# reset barrier
|
||||
try:
|
||||
if os.path.exists(BARRIER_FILE):
|
||||
os.remove(BARRIER_FILE)
|
||||
except:
|
||||
pass
|
||||
|
||||
print("Barrier 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("=" * 60)
|
||||
print(f"Rank {rank}/{world_size} START {datetime.now()}")
|
||||
print(f"Hostname: {socket.gethostname()}")
|
||||
print(f"{'='*60}")
|
||||
print("=" * 60)
|
||||
|
||||
# STEP 1: Signal that this worker is ready
|
||||
print(f"[{rank}] Signaling ready state...")
|
||||
max_wait = 300 # 5 minutes
|
||||
start_wait = time.time()
|
||||
# =========================
|
||||
# STEP 1: SAFE BARRIER
|
||||
# =========================
|
||||
print(f"[{rank}] entering barrier sync...")
|
||||
|
||||
while time.time() - start_wait < max_wait:
|
||||
with open(BARRIER_FILE, "a+") as f:
|
||||
f.write(f"{rank}\n")
|
||||
|
||||
# wait for all ranks
|
||||
while True:
|
||||
try:
|
||||
sync_state = json.loads(Variable.get('ddp_sync_state', default_var='{}'))
|
||||
with open(BARRIER_FILE, "r") as f:
|
||||
ready = set(f.read().strip().splitlines())
|
||||
|
||||
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']}")
|
||||
print(f"[{rank}] ready workers: {ready}")
|
||||
|
||||
# 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...")
|
||||
if len(ready) == world_size:
|
||||
print(f"[{rank}] ALL WORKERS READY")
|
||||
break
|
||||
|
||||
print(f"[{rank}] Waiting for other workers... ({len(sync_state.get('ready_workers', []))}/{world_size} ready)")
|
||||
time.sleep(2)
|
||||
except FileNotFoundError:
|
||||
pass
|
||||
|
||||
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!")
|
||||
time.sleep(1)
|
||||
|
||||
# Small delay to ensure all workers see the ready state
|
||||
time.sleep(3)
|
||||
time.sleep(2) # small stabilization delay
|
||||
|
||||
# STEP 3: Configure distributed environment
|
||||
# =========================
|
||||
# STEP 2: NCCL ENV
|
||||
# =========================
|
||||
os.environ['MASTER_ADDR'] = MASTER_ADDR
|
||||
os.environ['MASTER_PORT'] = MASTER_PORT
|
||||
os.environ['WORLD_SIZE'] = str(world_size)
|
||||
os.environ['RANK'] = str(rank)
|
||||
os.environ['NCCL_SOCKET_IFNAME'] = 'eth0'
|
||||
|
||||
# IMPORTANT FIX (IB stability toggle)
|
||||
os.environ["NCCL_IB_DISABLE"] = "1" # <- FIX #2 (can turn OFF later)
|
||||
os.environ["NCCL_SOCKET_IFNAME"] = "eth0"
|
||||
|
||||
os.environ['NCCL_DEBUG'] = 'INFO'
|
||||
os.environ['NCCL_TIMEOUT'] = str(NCCL_TIMEOUT)
|
||||
os.environ['NCCL_BLOCKING_WAIT'] = '1'
|
||||
|
||||
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}")
|
||||
print(f"[{rank}] env ready")
|
||||
|
||||
# STEP 4: Initialize process group
|
||||
print(f"[{rank}] Initializing process group (backend=nccl)...")
|
||||
try:
|
||||
# =========================
|
||||
# STEP 3: INIT PROCESS GROUP
|
||||
# =========================
|
||||
dist.init_process_group(
|
||||
backend="nccl",
|
||||
init_method="env://",
|
||||
@@ -104,28 +101,18 @@ def run_training_node_func(rank, world_size):
|
||||
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: {str(e)}")
|
||||
raise
|
||||
|
||||
# STEP 5: Setup GPU
|
||||
print(f"[{rank}] process group OK")
|
||||
|
||||
# =========================
|
||||
# STEP 4: 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)
|
||||
|
||||
@@ -136,138 +123,98 @@ def run_training_node_func(rank, world_size):
|
||||
)
|
||||
|
||||
criterion = nn.MSELoss()
|
||||
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001, momentum=0.9)
|
||||
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
|
||||
|
||||
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):
|
||||
# =========================
|
||||
# STEP 5: TRAIN LOOP
|
||||
# =========================
|
||||
for epoch in range(5):
|
||||
ddp_model.train()
|
||||
epoch_loss = 0.0
|
||||
loss_sum = 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)
|
||||
for _ in range(5):
|
||||
x = torch.randn(32, 10).to(device)
|
||||
y = torch.randn(32, 10).to(device)
|
||||
|
||||
optimizer.zero_grad()
|
||||
outputs = ddp_model(inputs)
|
||||
loss = criterion(outputs, labels)
|
||||
out = ddp_model(x)
|
||||
loss = criterion(out, y)
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
|
||||
epoch_loss += loss.item()
|
||||
loss_sum += loss.item()
|
||||
|
||||
avg_loss = epoch_loss / num_batches
|
||||
print(f"[{rank}] epoch {epoch} loss {loss_sum/5:.4f}")
|
||||
|
||||
# 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'
|
||||
}
|
||||
print(f"[{rank}] DONE")
|
||||
|
||||
return {"rank": rank, "status": "success"}
|
||||
|
||||
|
||||
def cleanup_sync_state_func():
|
||||
"""Clean up synchronization state"""
|
||||
try:
|
||||
Variable.delete('ddp_sync_state')
|
||||
print("Sync state cleaned up")
|
||||
if os.path.exists(BARRIER_FILE):
|
||||
os.remove(BARRIER_FILE)
|
||||
except:
|
||||
pass
|
||||
print("Barrier cleaned")
|
||||
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}")
|
||||
print("\n=== SUMMARY ===")
|
||||
|
||||
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}")
|
||||
res = ti.xcom_pull(task_ids=f"train_rank_{rank}")
|
||||
print(f"Rank {rank}: {res}")
|
||||
|
||||
print(f"\n✓ Training completed!")
|
||||
return {'status': 'success', 'workers': WORLD_SIZE}
|
||||
return {"status": "done"}
|
||||
|
||||
|
||||
# =========================
|
||||
# DAG
|
||||
# =========================
|
||||
with DAG(
|
||||
dag_id='pytorch_distributed_training_ddp_production',
|
||||
dag_id='pytorch_ddp_airflow_fixed',
|
||||
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,
|
||||
tags=['gpu', 'ddp', 'fixed'],
|
||||
) as dag:
|
||||
|
||||
# Preparation task
|
||||
prep = PythonOperator(
|
||||
task_id='prepare_training',
|
||||
python_callable=prepare_training_func,
|
||||
queue='gpu'
|
||||
queue='gpu',
|
||||
)
|
||||
|
||||
# Training tasks - one per rank
|
||||
training_tasks = []
|
||||
for i in range(WORLD_SIZE):
|
||||
task = PythonOperator(
|
||||
t = 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)
|
||||
training_tasks.append(t)
|
||||
|
||||
# Cleanup task
|
||||
cleanup = PythonOperator(
|
||||
task_id='cleanup_sync_state',
|
||||
task_id='cleanup',
|
||||
python_callable=cleanup_sync_state_func,
|
||||
queue='gpu',
|
||||
trigger_rule='all_done'
|
||||
trigger_rule='all_done',
|
||||
)
|
||||
|
||||
# Summary task
|
||||
summary = PythonOperator(
|
||||
task_id='training_summary',
|
||||
task_id='summary',
|
||||
python_callable=training_summary_func,
|
||||
queue='gpu',
|
||||
trigger_rule='all_done'
|
||||
trigger_rule='all_done',
|
||||
)
|
||||
|
||||
# Set dependencies
|
||||
prep >> training_tasks >> cleanup >> summary
|
||||
|
||||
|
||||
Reference in New Issue
Block a user