374 lines
13 KiB
Python
374 lines
13 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
|
|
WORKER_PORT = 12345
|
|
# Cluster spec will be built dynamically based on actual worker hostnames
|
|
|
|
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 = {
|
|
'worker_addresses': {}, # {task_index: 'hostname:port'}
|
|
'ready_workers': [],
|
|
'completed_workers': [], # Track workers that finished training
|
|
'training_started': False,
|
|
'start_time': None
|
|
}
|
|
Variable.set('tf_sync_state', json.dumps(sync_state))
|
|
print("Training preparation complete. Sync state initialized.")
|
|
return True
|
|
|
|
|
|
def run_training_node_func(task_index):
|
|
"""Execute TensorFlow distributed training"""
|
|
import os
|
|
import socket
|
|
from datetime import datetime
|
|
import time
|
|
|
|
hostname = socket.gethostname()
|
|
my_address = f"{hostname}.airflow-worker-gpu:{WORKER_PORT}"
|
|
|
|
print(f"{'='*60}")
|
|
print(f"Worker {task_index}/{WORLD_SIZE} - Starting at {datetime.now()}")
|
|
print(f"Hostname: {hostname}")
|
|
print(f"My address: {my_address}")
|
|
print(f"{'='*60}")
|
|
|
|
# STEP 1: Register worker address and wait for all workers
|
|
print(f"[Worker-{task_index}] Registering address and waiting for all workers...")
|
|
max_wait = 300 # 5 minutes
|
|
start_wait = time.time()
|
|
cluster_spec = None
|
|
|
|
while time.time() - start_wait < max_wait:
|
|
try:
|
|
sync_state = json.loads(Variable.get('tf_sync_state', default_var='{}'))
|
|
worker_addresses = sync_state.setdefault('worker_addresses', {})
|
|
|
|
# Register my address
|
|
if str(task_index) not in worker_addresses:
|
|
worker_addresses[str(task_index)] = my_address
|
|
sync_state['worker_addresses'] = worker_addresses
|
|
Variable.set('tf_sync_state', json.dumps(sync_state))
|
|
print(f"[Worker-{task_index}] Registered address: {my_address}")
|
|
|
|
# Mark as ready
|
|
if task_index not in sync_state.get('ready_workers', []):
|
|
sync_state.setdefault('ready_workers', []).append(task_index)
|
|
Variable.set('tf_sync_state', json.dumps(sync_state))
|
|
|
|
# STEP 2: Wait for all workers to register
|
|
if len(worker_addresses) == WORLD_SIZE:
|
|
print(f"[Worker-{task_index}] All {WORLD_SIZE} workers registered!")
|
|
# Build cluster spec from registered addresses
|
|
worker_list = [worker_addresses[str(i)] for i in range(WORLD_SIZE)]
|
|
cluster_spec = {'worker': worker_list}
|
|
print(f"[Worker-{task_index}] Cluster spec: {cluster_spec}")
|
|
break
|
|
|
|
print(f"[Worker-{task_index}] Waiting for workers... ({len(worker_addresses)}/{WORLD_SIZE} registered)")
|
|
time.sleep(2)
|
|
|
|
except Exception as e:
|
|
print(f"[Worker-{task_index}] Error during sync: {e}")
|
|
time.sleep(2)
|
|
else:
|
|
raise RuntimeError(f"[Worker-{task_index}] Timeout waiting for all workers to register!")
|
|
|
|
if cluster_spec is None:
|
|
raise RuntimeError(f"[Worker-{task_index}] Failed to build cluster spec!")
|
|
|
|
# Small delay to ensure all workers see the complete cluster
|
|
time.sleep(3)
|
|
|
|
# STEP 3: Configure TF_CONFIG BEFORE importing TensorFlow (CRITICAL!)
|
|
tf_config = {
|
|
'cluster': cluster_spec,
|
|
'task': {
|
|
'type': 'worker',
|
|
'index': task_index
|
|
}
|
|
}
|
|
|
|
os.environ['TF_CONFIG'] = json.dumps(tf_config)
|
|
|
|
print(f"[Worker-{task_index}] TF_CONFIG set:")
|
|
print(f" Cluster: {cluster_spec}")
|
|
print(f" Task type: worker")
|
|
print(f" Task index: {task_index}")
|
|
|
|
# NOW import TensorFlow after TF_CONFIG is set
|
|
import tensorflow as tf
|
|
print(f"TensorFlow version: {tf.__version__}")
|
|
|
|
# STEP 4: Initialize distribution strategy IMMEDIATELY (CRITICAL!)
|
|
# Must happen before any other TensorFlow operations
|
|
print(f"[Worker-{task_index}] Initializing MultiWorkerMirroredStrategy...")
|
|
|
|
# Configure communication options for NCCL
|
|
communication_options = tf.distribute.experimental.CommunicationOptions(
|
|
implementation=tf.distribute.experimental.CommunicationImplementation.NCCL
|
|
)
|
|
|
|
strategy = tf.distribute.MultiWorkerMirroredStrategy(
|
|
communication_options=communication_options
|
|
)
|
|
|
|
print(f"[Worker-{task_index}] ✓ Strategy initialized!")
|
|
print(f" Number of devices in sync: {strategy.num_replicas_in_sync}")
|
|
|
|
# STEP 5: GPU info (after strategy is initialized)
|
|
gpus = tf.config.list_physical_devices('GPU')
|
|
print(f"[Worker-{task_index}] Available GPUs: {len(gpus)}")
|
|
if gpus:
|
|
print(f"[Worker-{task_index}] GPU: {gpus[0].name}")
|
|
|
|
# STEP 6: Define model and training within strategy scope
|
|
with strategy.scope():
|
|
# Create model
|
|
model = tf.keras.Sequential([
|
|
tf.keras.layers.Dense(128, activation='relu', input_shape=(10,)),
|
|
tf.keras.layers.Dense(128, activation='relu'),
|
|
tf.keras.layers.Dense(10)
|
|
])
|
|
|
|
# Compile model
|
|
model.compile(
|
|
optimizer=tf.keras.optimizers.SGD(learning_rate=0.001, momentum=0.9),
|
|
loss=tf.keras.losses.MeanSquaredError(),
|
|
metrics=['mae']
|
|
)
|
|
|
|
print(f"[Worker-{task_index}] Model created with {model.count_params()} parameters")
|
|
|
|
# STEP 7: Create synthetic dataset with manual sharding
|
|
print(f"[Worker-{task_index}] Creating dataset...")
|
|
|
|
import numpy as np
|
|
|
|
batch_size = 32
|
|
total_samples = 160
|
|
steps_per_epoch = 5
|
|
num_workers = strategy.num_replicas_in_sync
|
|
|
|
# Create synthetic data - same for all workers initially
|
|
np.random.seed(42)
|
|
x_data = np.random.randn(total_samples, 10).astype(np.float32)
|
|
y_data = np.random.randn(total_samples, 10).astype(np.float32)
|
|
|
|
# Manually shard data for this worker
|
|
# Each worker gets a disjoint subset
|
|
samples_per_worker = total_samples // num_workers
|
|
start_idx = task_index * samples_per_worker
|
|
end_idx = start_idx + samples_per_worker
|
|
|
|
x_worker = x_data[start_idx:end_idx]
|
|
y_worker = y_data[start_idx:end_idx]
|
|
|
|
print(f"[Worker-{task_index}] Worker data: indices {start_idx}:{end_idx} ({len(x_worker)} samples)")
|
|
|
|
# Create dataset for this worker
|
|
dataset = tf.data.Dataset.from_tensor_slices((x_worker, y_worker))
|
|
dataset = dataset.shuffle(len(x_worker))
|
|
dataset = dataset.batch(batch_size, drop_remainder=True)
|
|
dataset = dataset.repeat()
|
|
|
|
# Disable auto-sharding since we manually sharded
|
|
options = tf.data.Options()
|
|
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
|
|
dataset = dataset.with_options(options)
|
|
|
|
print(f"[Worker-{task_index}] Dataset created (batch size: {batch_size})")
|
|
|
|
# STEP 8: Custom training loop (avoids PerReplica issues)
|
|
print(f"\n[Worker-{task_index}] {'='*60}")
|
|
print(f"[Worker-{task_index}] Starting Training Loop")
|
|
print(f"[Worker-{task_index}] {'='*60}")
|
|
|
|
num_epochs = 10
|
|
num_batches = 2 # Each worker will train on 2 batches per epoch
|
|
|
|
@tf.function
|
|
def train_step(x, y):
|
|
with tf.GradientTape() as tape:
|
|
predictions = model(x, training=True)
|
|
loss = model.compiled_loss(y, predictions)
|
|
|
|
gradients = tape.gradient(loss, model.trainable_variables)
|
|
model.optimizer.apply_gradients(zip(gradients, model.trainable_variables))
|
|
|
|
return loss
|
|
|
|
history_loss = []
|
|
|
|
for epoch in range(num_epochs):
|
|
epoch_losses = []
|
|
|
|
# Get batches from dataset
|
|
batch_iter = iter(dataset)
|
|
for batch_idx in range(num_batches):
|
|
x_batch, y_batch = next(batch_iter)
|
|
|
|
# Distributed training step
|
|
per_replica_loss = strategy.run(train_step, args=(x_batch, y_batch))
|
|
|
|
# Reduce loss across replicas
|
|
loss = strategy.reduce(tf.distribute.ReduceOp.MEAN, per_replica_loss, axis=None)
|
|
epoch_losses.append(float(loss.numpy()))
|
|
|
|
avg_loss = sum(epoch_losses) / len(epoch_losses)
|
|
history_loss.append(avg_loss)
|
|
|
|
if task_index == 0:
|
|
print(f"[Worker-{task_index}] Epoch {epoch+1}/{num_epochs} | Loss: {avg_loss:.6f}")
|
|
|
|
print(f"\n[Worker-{task_index}] {'='*60}")
|
|
print(f"[Worker-{task_index}] Training Complete!")
|
|
print(f"[Worker-{task_index}] {'='*60}")
|
|
|
|
# Get final metrics
|
|
final_loss = history_loss[-1]
|
|
|
|
if task_index == 0:
|
|
print(f"[Worker-{task_index}] Final Loss: {final_loss:.6f}")
|
|
|
|
# STEP 9: Synchronization barrier - wait for all workers to finish training
|
|
print(f"[Worker-{task_index}] Waiting at barrier for all workers to complete...")
|
|
import time
|
|
max_wait = 120
|
|
start_wait = time.time()
|
|
|
|
while time.time() - start_wait < max_wait:
|
|
try:
|
|
sync_state = json.loads(Variable.get('tf_sync_state', default_var='{}'))
|
|
completed_workers = sync_state.get('completed_workers', [])
|
|
|
|
if task_index not in completed_workers:
|
|
completed_workers.append(task_index)
|
|
sync_state['completed_workers'] = completed_workers
|
|
Variable.set('tf_sync_state', json.dumps(sync_state))
|
|
print(f"[Worker-{task_index}] Marked as completed ({len(completed_workers)}/{WORLD_SIZE})")
|
|
|
|
if len(completed_workers) == WORLD_SIZE:
|
|
print(f"[Worker-{task_index}] All workers completed! Proceeding to cleanup.")
|
|
break
|
|
|
|
time.sleep(2)
|
|
except Exception as e:
|
|
print(f"[Worker-{task_index}] Error during completion sync: {e}")
|
|
time.sleep(2)
|
|
else:
|
|
print(f"[Worker-{task_index}] Warning: Timeout at completion barrier")
|
|
|
|
time.sleep(2) # Small delay to ensure all workers pass barrier
|
|
|
|
# STEP 10: Save model (only chief worker)
|
|
if task_index == 0:
|
|
model_path = '/tmp/tf_distributed_model.keras'
|
|
print(f"[Worker-{task_index}] Saving model to {model_path}...")
|
|
model.save(model_path)
|
|
print(f"[Worker-{task_index}] ✓ Model saved successfully!")
|
|
|
|
print(f"[Worker-{task_index}] Exiting gracefully...")
|
|
|
|
return {
|
|
'worker_index': task_index,
|
|
'final_loss': float(final_loss),
|
|
'epochs_completed': num_epochs,
|
|
'status': 'success'
|
|
}
|
|
|
|
|
|
def cleanup_sync_state_func():
|
|
"""Clean up synchronization state"""
|
|
try:
|
|
Variable.delete('tf_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"TENSORFLOW DISTRIBUTED TRAINING SUMMARY")
|
|
print(f"{'='*60}")
|
|
|
|
for worker_idx in range(WORLD_SIZE):
|
|
try:
|
|
result = ti.xcom_pull(task_ids=f"train_worker_{worker_idx}")
|
|
if result and result.get('status') == 'success':
|
|
print(f"Worker {result['worker_index']}: ✓ 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 worker {worker_idx}: {e}")
|
|
|
|
print(f"\n✓ Distributed training completed across {WORLD_SIZE} workers!")
|
|
return {'status': 'success', 'workers': WORLD_SIZE}
|
|
|
|
|
|
with DAG(
|
|
dag_id='tensorflow_distributed_training_multiworker',
|
|
default_args=default_args,
|
|
schedule=None,
|
|
start_date=pendulum.today('UTC').add(days=-1),
|
|
catchup=False,
|
|
tags=['gpu', 'ml', 'distributed', 'tensorflow'],
|
|
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 worker
|
|
training_tasks = []
|
|
for i in range(WORLD_SIZE):
|
|
task = PythonOperator(
|
|
task_id=f'train_worker_{i}',
|
|
python_callable=run_training_node_func,
|
|
op_kwargs={'task_index': i},
|
|
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
|