e61e8f54da
Changes: - CLAUDE.md: Update OOM fix validation status - Add comprehensive documentation (30+ markdown reports) - LSTM encoder varmap bug fix (tft/lstm_encoder.rs:290) - Quantized LSTM layer matching fix (tft/quantized_lstm.rs) - Hyperopt paths module (ml/src/hyperopt/paths.rs) - Training path tests for all adapters (DQN, MAMBA-2, PPO, TFT) - Checkpoint integrity tests - Script cleanup: Remove 29 obsolete deployment scripts - Archive old scripts to scripts/archive/ - New deployment utilities: check_gpu_availability.py, monitor_hyperopt.sh Validation: - OOM fixes validated: 5/5 trials successful (pod b6kc3mc5lbjiro) - Batch-size-max 256 tested successfully - All hyperopt adapters working correctly 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
420 lines
15 KiB
Python
Executable File
420 lines
15 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""
|
|
RunPod Deployment Script - FIXED VERSION
|
|
Scans for best value GPU in EUR-IS region and deploys a pod on SECURE cloud.
|
|
|
|
KEY FIX: Uses REST API for availability-aware deployment instead of GraphQL global counts.
|
|
"""
|
|
|
|
import os
|
|
import sys
|
|
import argparse
|
|
import requests
|
|
from dotenv import load_dotenv
|
|
|
|
# Load environment variables from .env.runpod
|
|
env_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), '.env.runpod')
|
|
load_dotenv(env_path)
|
|
|
|
RUNPOD_API_KEY = os.getenv('RUNPOD_API_KEY')
|
|
RUNPOD_VOLUME_ID = os.getenv('RUNPOD_VOLUME_ID')
|
|
RUNPOD_CONTAINER_REGISTRY_AUTH_ID = os.getenv('RUNPOD_CONTAINER_REGISTRY_AUTH_ID')
|
|
|
|
if not RUNPOD_API_KEY:
|
|
print("ERROR: RUNPOD_API_KEY not found in .env.runpod")
|
|
sys.exit(1)
|
|
|
|
if not RUNPOD_VOLUME_ID:
|
|
print("ERROR: RUNPOD_VOLUME_ID not found in .env.runpod")
|
|
sys.exit(1)
|
|
|
|
# EUR-IS datacenters to scan
|
|
# CRITICAL FIX: Volume se3zdnb5o4 is in EUR-IS-1 ONLY!
|
|
# If pod deploys to EUR-IS-2 or EUR-IS-3, volume will NOT be accessible
|
|
EUR_IS_DATACENTERS = ['EUR-IS-1'] # ONLY EUR-IS-1 for volume mounting!
|
|
|
|
# REST API endpoint (NEW - supports datacenter filtering)
|
|
REST_API_URL = "https://rest.runpod.io/v1/pods"
|
|
|
|
# GraphQL endpoint (still needed for listing GPU types and pricing)
|
|
GRAPHQL_ENDPOINT = "https://api.runpod.io/graphql"
|
|
|
|
# GraphQL query to fetch GPU types with global availability and pricing
|
|
GPU_QUERY = """
|
|
{
|
|
gpuTypes {
|
|
id
|
|
displayName
|
|
memoryInGb
|
|
secureCloud
|
|
communityCloud
|
|
lowestPrice(input: {gpuCount: 1}) {
|
|
uninterruptablePrice
|
|
}
|
|
}
|
|
}
|
|
"""
|
|
|
|
def query_graphql(query, variables=None):
|
|
"""Execute GraphQL query against RunPod API."""
|
|
headers = {
|
|
"Content-Type": "application/json",
|
|
"Authorization": f"Bearer {RUNPOD_API_KEY}"
|
|
}
|
|
|
|
payload = {"query": query}
|
|
if variables:
|
|
payload["variables"] = variables
|
|
|
|
try:
|
|
response = requests.post(GRAPHQL_ENDPOINT, json=payload, headers=headers, timeout=30)
|
|
response.raise_for_status()
|
|
result = response.json()
|
|
|
|
# Check for GraphQL errors
|
|
if 'errors' in result:
|
|
error_msg = result['errors'][0].get('message', 'Unknown error')
|
|
print(f"ERROR: GraphQL errors: {result['errors']}")
|
|
return None
|
|
|
|
return result
|
|
except requests.exceptions.RequestException as e:
|
|
print(f"ERROR: Failed to query RunPod GraphQL API: {e}")
|
|
return None
|
|
|
|
|
|
def get_available_gpu_types():
|
|
"""
|
|
Query available GPU types with ≥16GB VRAM that have SOME availability in SECURE cloud.
|
|
|
|
NOTE: This returns GLOBAL secure cloud availability, not EUR-IS specific.
|
|
The actual datacenter filtering happens during deployment via REST API.
|
|
"""
|
|
print(" Querying GPU types and pricing...")
|
|
|
|
data = query_graphql(GPU_QUERY)
|
|
if not data:
|
|
return []
|
|
|
|
gpu_types = data.get('data', {}).get('gpuTypes', [])
|
|
|
|
# Filter criteria:
|
|
# 1. memoryInGb >= 16
|
|
# 2. secureCloud > 0 (available SOMEWHERE in secure cloud - not necessarily EUR-IS)
|
|
# 3. Has pricing information
|
|
available_gpus = []
|
|
for gpu in gpu_types:
|
|
memory = gpu.get('memoryInGb', 0)
|
|
secure_count = gpu.get('secureCloud', 0)
|
|
lowest_price = gpu.get('lowestPrice', {})
|
|
price = lowest_price.get('uninterruptablePrice') if lowest_price else None
|
|
|
|
if memory >= 16 and secure_count > 0 and price is not None:
|
|
available_gpus.append({
|
|
'id': gpu.get('id', ''),
|
|
'name': gpu.get('displayName', 'Unknown'),
|
|
'vram': memory,
|
|
'price': float(price),
|
|
'global_available': secure_count # This is GLOBAL, not EUR-IS specific
|
|
})
|
|
|
|
if available_gpus:
|
|
# Sort by price (cheapest first)
|
|
available_gpus.sort(key=lambda x: x['price'])
|
|
print(f" ✅ Found {len(available_gpus)} GPU type(s) with global secure cloud availability")
|
|
else:
|
|
print(f" ⚠️ No GPU types found with ≥16GB VRAM in secure cloud")
|
|
|
|
return available_gpus
|
|
|
|
|
|
def deploy_pod_rest_api(gpu, image, command, container_disk, datacenters, dry_run=False):
|
|
"""
|
|
Deploy a pod using the REST API with datacenter-specific availability filtering.
|
|
|
|
This is the KEY FIX: REST API checks EUR-IS datacenter availability at deployment time,
|
|
not relying on global GraphQL counts.
|
|
"""
|
|
pod_name = "foxhunt-training"
|
|
|
|
# Build REST API request payload
|
|
# CRITICAL: Only use fields documented in official RunPod REST API
|
|
# Reference: https://docs.runpod.io/api-reference/pods/POST/pods
|
|
deployment_payload = {
|
|
"cloudType": "SECURE", # CRITICAL: Only secure cloud
|
|
"computeType": "GPU", # GPU pod (required field)
|
|
"dataCenterIds": datacenters, # CRITICAL: EUR-IS specific datacenters
|
|
"dataCenterPriority": "availability", # Try datacenters in order, prefer available
|
|
"gpuTypeIds": [gpu['id']], # GPU type to deploy
|
|
"gpuTypePriority": "availability", # Use available GPU
|
|
"gpuCount": 1,
|
|
"name": pod_name,
|
|
"imageName": image,
|
|
"containerDiskInGb": container_disk,
|
|
"volumeInGb": 0, # Using network volume instead
|
|
"networkVolumeId": RUNPOD_VOLUME_ID,
|
|
"volumeMountPath": "/runpod-volume", # CRITICAL: WHERE to mount the volume
|
|
"ports": ["8888/http", "22/tcp"],
|
|
"env": {},
|
|
"interruptible": False, # On-demand (non-spot)
|
|
"minRAMPerGPU": 8,
|
|
"minVCPUPerGPU": 2
|
|
}
|
|
|
|
# Add Docker command if specified
|
|
# RunPod expects dockerStartCmd as an array of strings (shell arguments)
|
|
# Split the command string into proper arguments
|
|
if command:
|
|
# Split command into arguments (respects quoted strings)
|
|
import shlex
|
|
deployment_payload["dockerStartCmd"] = shlex.split(command)
|
|
|
|
# Add Docker registry authentication if configured
|
|
if RUNPOD_CONTAINER_REGISTRY_AUTH_ID:
|
|
deployment_payload["containerRegistryAuthId"] = RUNPOD_CONTAINER_REGISTRY_AUTH_ID
|
|
|
|
print("\n" + "="*70)
|
|
print("DEPLOYMENT PLAN")
|
|
print("="*70)
|
|
print(f"Pod Name: {pod_name}")
|
|
print(f"GPU: {gpu['name']} ({gpu['vram']}GB VRAM)")
|
|
print(f"Datacenters: {', '.join(datacenters)} (tries in order)")
|
|
print(f"Price: ${gpu['price']:.3f}/hr (estimate)")
|
|
print(f"Docker Image: {image}")
|
|
print(f"Container Disk: {container_disk}GB")
|
|
print(f"Network Volume: {RUNPOD_VOLUME_ID} → /runpod-volume")
|
|
print(f"Ports: 8888/http (Jupyter), 22/tcp (SSH)")
|
|
print(f"Registry Auth: {RUNPOD_CONTAINER_REGISTRY_AUTH_ID}")
|
|
if command:
|
|
print(f"Command: {command}")
|
|
print("="*70)
|
|
|
|
if dry_run:
|
|
print("\nDRY RUN: Skipping actual deployment")
|
|
print("\nPayload would be:")
|
|
import json
|
|
print(json.dumps(deployment_payload, indent=2))
|
|
return None
|
|
|
|
print("\nDeploying pod via REST API (checks EUR-IS availability)...")
|
|
|
|
headers = {
|
|
"Content-Type": "application/json",
|
|
"Authorization": f"Bearer {RUNPOD_API_KEY}"
|
|
}
|
|
|
|
try:
|
|
response = requests.post(REST_API_URL, json=deployment_payload, headers=headers, timeout=60)
|
|
|
|
# Debug output
|
|
print(f" HTTP Status: {response.status_code}")
|
|
|
|
# SUCCESS: HTTP 200 (OK) or HTTP 201 (Created)
|
|
if response.status_code in [200, 201]:
|
|
try:
|
|
pod_data = response.json()
|
|
|
|
# Validate pod_data structure
|
|
if not isinstance(pod_data, dict):
|
|
print(f" ⚠️ Unexpected response format: {type(pod_data)}")
|
|
print(f" Raw response: {response.text[:500]}")
|
|
return None
|
|
|
|
# Check if pod was actually created
|
|
if 'id' not in pod_data:
|
|
print(f" ⚠️ Pod data missing 'id' field")
|
|
print(f" Raw response: {response.text[:500]}")
|
|
return None
|
|
|
|
print(f" ✅ Pod created successfully! ID: {pod_data['id']}")
|
|
return pod_data
|
|
except ValueError as e:
|
|
print(f" ⚠️ Failed to parse JSON response: {e}")
|
|
print(f" Raw response: {response.text[:500]}")
|
|
return None
|
|
elif response.status_code == 400:
|
|
try:
|
|
error_data = response.json()
|
|
error_msg = error_data.get('error', 'Unknown error')
|
|
print(f" ⚠️ Deployment failed: {error_msg}")
|
|
|
|
# Check if it's an availability issue
|
|
if 'not available' in error_msg.lower() or 'no machines' in error_msg.lower():
|
|
print(f" 💡 GPU not available in EUR-IS datacenters at this time")
|
|
except ValueError:
|
|
print(f" ⚠️ Deployment failed: {response.text[:200]}")
|
|
|
|
return None
|
|
else:
|
|
print(f" ❌ Unexpected response (status {response.status_code}): {response.text[:200]}")
|
|
return None
|
|
|
|
except requests.exceptions.RequestException as e:
|
|
print(f" ⚠️ Deployment failed: {str(e)[:100]}")
|
|
return None
|
|
|
|
|
|
def display_deployment_result(pod_data):
|
|
"""Display deployment results and connection info."""
|
|
print("\n" + "="*70)
|
|
print("✅ POD DEPLOYED SUCCESSFULLY")
|
|
print("="*70)
|
|
print(f"Pod ID: {pod_data['id']}")
|
|
|
|
# Extract GPU info safely
|
|
machine = pod_data.get('machine', {})
|
|
gpu_info = machine.get('gpuType', {}) if machine else {}
|
|
print(f"GPU: {gpu_info.get('displayName', 'N/A')}")
|
|
|
|
print(f"GPU Count: {pod_data.get('gpu', {}).get('count', 'N/A')}")
|
|
print(f"Cost: ${pod_data.get('costPerHr', 'N/A')}/hr")
|
|
|
|
# Extract datacenter
|
|
datacenter = machine.get('dataCenterId', 'N/A')
|
|
print(f"Datacenter: {datacenter}")
|
|
|
|
print(f"Image: {pod_data.get('imageName', pod_data.get('image', 'N/A'))}")
|
|
print(f"Container Disk: {pod_data['containerDiskInGb']}GB")
|
|
print(f"Status: {pod_data.get('desiredStatus', 'RUNNING')}")
|
|
print("="*70)
|
|
print("\n📝 NEXT STEPS:")
|
|
print("1. Wait 2-3 minutes for pod to initialize")
|
|
print(f"2. Access Jupyter at: https://{pod_data['id']}-8888.proxy.runpod.net")
|
|
print(f"3. SSH access: ssh root@{pod_data['id']}.ssh.runpod.io")
|
|
print("4. Monitor pod: https://www.runpod.io/console/pods")
|
|
print(f"\n⚠️ IMPORTANT: Pod will cost ${pod_data.get('costPerHr', 'N/A')}/hr - remember to stop when done!")
|
|
print()
|
|
|
|
|
|
def main():
|
|
"""Main execution function."""
|
|
parser = argparse.ArgumentParser(
|
|
description="Deploy a RunPod GPU pod in EUR-IS region (SECURE cloud) - FIXED VERSION",
|
|
formatter_class=argparse.RawDescriptionHelpFormatter,
|
|
epilog="""
|
|
Examples:
|
|
# Auto-select best value GPU with default training command
|
|
./runpod_deploy.py
|
|
|
|
# Deploy with specific GPU
|
|
./runpod_deploy.py --gpu-type "RTX 4090"
|
|
|
|
# Custom training command (overrides Dockerfile CMD)
|
|
./runpod_deploy.py --command "--parquet-file /runpod-volume/test_data/ES_FUT_180d.parquet --epochs 100"
|
|
|
|
# Custom image and command
|
|
./runpod_deploy.py --image runpod/pytorch:2.1.0 --command "jupyter lab --ip=0.0.0.0 --allow-root"
|
|
|
|
# Dry run (show plan without deploying)
|
|
./runpod_deploy.py --dry-run
|
|
|
|
KEY FIXES:
|
|
- Uses REST API for deployment (checks EUR-IS datacenter availability)
|
|
- Properly formats dockerStartCmd as array (RunPod API requirement)
|
|
- Default command: TFT training with Parquet data (overrides Dockerfile CMD)
|
|
"""
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--gpu-type',
|
|
help='Preferred GPU type (e.g., "RTX 4090"). Auto-selects best value if not specified.'
|
|
)
|
|
parser.add_argument(
|
|
'--image',
|
|
default='jgrusewski/foxhunt:latest',
|
|
help='Docker image to use (default: jgrusewski/foxhunt:latest)'
|
|
)
|
|
parser.add_argument(
|
|
'--command',
|
|
default='--parquet-file /runpod-volume/test_data/ES_FUT_180d.parquet --epochs 50 --learning-rate 0.001',
|
|
help='Training command arguments (default: TFT training with Parquet data)'
|
|
)
|
|
parser.add_argument(
|
|
'--container-disk',
|
|
type=int,
|
|
default=50,
|
|
help='Container disk size in GB (default: 50)'
|
|
)
|
|
parser.add_argument(
|
|
'--dry-run',
|
|
action='store_true',
|
|
help='Show deployment plan without actually deploying'
|
|
)
|
|
|
|
args = parser.parse_args()
|
|
|
|
print("🔍 Querying available GPU types (global secure cloud)...")
|
|
|
|
# Query available GPUs (global availability)
|
|
gpus = get_available_gpu_types()
|
|
|
|
if not gpus:
|
|
print("\nERROR: No GPUs available with ≥16GB VRAM in SECURE cloud")
|
|
print("\n💡 TIP: This checks global availability. EUR-IS specific availability")
|
|
print(" is checked during deployment via REST API.")
|
|
sys.exit(1)
|
|
|
|
print(f"\n✅ Found {len(gpus)} GPU type(s) to try")
|
|
|
|
# Create priority list: user preference (if specified), then sorted by price
|
|
priority_gpus = []
|
|
|
|
# Add user's preferred GPU if specified
|
|
if args.gpu_type:
|
|
for gpu in gpus:
|
|
if args.gpu_type.lower() in gpu['name'].lower():
|
|
priority_gpus.append(gpu)
|
|
break
|
|
if not priority_gpus:
|
|
print(f"WARNING: Preferred GPU '{args.gpu_type}' not found, trying all GPUs...")
|
|
|
|
# Add remaining GPUs sorted by price (cheapest first)
|
|
# gpus list is already sorted by price at line 121
|
|
for gpu in gpus:
|
|
if gpu not in priority_gpus:
|
|
priority_gpus.append(gpu)
|
|
|
|
# Try each GPU until one succeeds
|
|
attempted_gpus = []
|
|
pod_data = None
|
|
|
|
for gpu in priority_gpus:
|
|
if gpu in attempted_gpus:
|
|
continue
|
|
|
|
attempted_gpus.append(gpu)
|
|
|
|
print(f"\n🎯 Attempting deployment: {gpu['name']} (${gpu['price']:.3f}/hr)...")
|
|
print(f" (Global availability: {gpu['global_available']} in secure cloud)")
|
|
print(f" (Will check EUR-IS specific availability during deployment...)")
|
|
|
|
pod_data = deploy_pod_rest_api(
|
|
gpu,
|
|
args.image,
|
|
args.command,
|
|
args.container_disk,
|
|
EUR_IS_DATACENTERS,
|
|
args.dry_run
|
|
)
|
|
|
|
if pod_data:
|
|
break
|
|
else:
|
|
print(f"❌ {gpu['name']} not available in EUR-IS, trying next option...")
|
|
|
|
# Display results
|
|
if pod_data:
|
|
display_deployment_result(pod_data)
|
|
else:
|
|
print("\n❌ ERROR: Failed to deploy pod on any available GPU in EUR-IS")
|
|
print("\nAttempted GPUs:")
|
|
for gpu in attempted_gpus:
|
|
print(f" - {gpu['name']} ({gpu['vram']}GB VRAM) @ ${gpu['price']:.3f}/hr")
|
|
print("\n💡 TIP: RunPod availability changes frequently. Try again in a few minutes.")
|
|
print(" The script now correctly checks EUR-IS datacenter availability,")
|
|
print(" not just global secure cloud counts.")
|
|
sys.exit(1)
|
|
|
|
if __name__ == "__main__":
|
|
main()
|