# Docker Compose for GPU Deployment - Medical GraphRAG Assistant
# Full stack: IRIS FHIR + NIM LLM + NV-CLIP on multi-GPU instances
#
# Recommended: g5.12xlarge (4x A10G, 96GB VRAM) or g5.xlarge (1x A10G, 24GB)
#
# Usage:
# # Set your NGC API key
# export NGC_API_KEY=nvapi-xxx
#
# # Start everything
# docker compose -f docker-compose.gpu.yml up -d
#
# # Check status
# docker compose -f docker-compose.gpu.yml ps
#
# # View logs
# docker compose -f docker-compose.gpu.yml logs -f
#
# # Stop everything
# docker compose -f docker-compose.gpu.yml down
services:
# ============================================================================
# IRIS for Health - FHIR Repository + Vector Database
# Uses custom-built image with FHIR packages pre-installed
# ============================================================================
iris-fhir:
build:
context: ./Dockerfhir
dockerfile: Dockerfile
image: dockerfhir-iris-fhir:latest
container_name: iris-fhir
ports:
- "32782:1972" # SuperServer (SQL)
- "32783:52773" # Web Portal + FHIR
volumes:
- iris-data:/durable
# IRIS Community Edition is limited to ~8 cores
# g5.12xlarge has 48 vCPUs - must limit to avoid license error
deploy:
resources:
limits:
cpus: '4'
restart: unless-stopped
healthcheck:
# Use wget since curl not in container
test: ["CMD-SHELL", "wget -q --spider http://localhost:52773/csp/sys/UtilHome.csp || exit 1"]
interval: 30s
timeout: 10s
retries: 5
start_period: 180s
# ============================================================================
# NIM LLM - Llama 3.1 8B for text generation and entity extraction
# Uses GPUs 0-2 on multi-GPU instances (tensor parallel)
# ============================================================================
nim-llm:
image: nvcr.io/nim/meta/llama-3.1-8b-instruct:latest
container_name: nim-llm
ports:
- "8001:8000"
environment:
- NGC_API_KEY=${NGC_API_KEY}
- CUDA_VISIBLE_DEVICES=0,1,2
deploy:
resources:
reservations:
devices:
- driver: nvidia
capabilities: [gpu]
restart: unless-stopped
healthcheck:
test: ["CMD", "curl", "-f", "http://localhost:8000/v1/models"]
interval: 30s
timeout: 10s
retries: 10
start_period: 180s
depends_on:
iris-fhir:
condition: service_healthy
# ============================================================================
# NV-CLIP - Multimodal embeddings for medical image search
# Uses GPU 3 on multi-GPU instances (dedicated)
# ============================================================================
nim-nvclip:
image: nvcr.io/nim/nvidia/nvclip:2.0.0
container_name: nim-nvclip
ports:
- "8002:8000"
environment:
- NGC_API_KEY=${NGC_API_KEY}
- CUDA_VISIBLE_DEVICES=3
deploy:
resources:
reservations:
devices:
- driver: nvidia
capabilities: [gpu]
restart: unless-stopped
healthcheck:
test: ["CMD", "curl", "-sf", "-X", "POST", "http://localhost:8000/v1/embeddings", "-H", "Content-Type: application/json", "-d", "{\"input\":[\"test\"],\"model\":\"nvidia/nvclip\"}"]
interval: 30s
timeout: 15s
retries: 10
start_period: 180s
depends_on:
iris-fhir:
condition: service_healthy
# ============================================================================
# Streamlit UI - Chat interface for medical search
# ============================================================================
streamlit:
build:
context: .
dockerfile: Dockerfile.streamlit
container_name: streamlit-ui
ports:
- "8501:8501"
environment:
- IRIS_HOST=iris-fhir
- IRIS_PORT=1972
- IRIS_NAMESPACE=%SYS
- IRIS_USERNAME=_SYSTEM
- IRIS_PASSWORD=SYS
- FHIR_BASE_URL=http://iris-fhir:52773/csp/healthshare/demo/fhir/r4
- FHIR_USERNAME=_SYSTEM
- FHIR_PASSWORD=SYS
- NIM_LLM_URL=http://nim-llm:8000/v1
- NVCLIP_BASE_URL=http://nim-nvclip:8000/v1
- NVIDIA_API_KEY=${NGC_API_KEY}
restart: unless-stopped
depends_on:
nim-llm:
condition: service_healthy
nim-nvclip:
condition: service_healthy
profiles:
- full # Only start with: docker compose --profile full up
volumes:
iris-data:
driver: local
# =============================================================================
# DEPLOYMENT PROFILES
# =============================================================================
#
# 1. FULL STACK (g5.12xlarge - 4 GPUs):
# docker compose -f docker-compose.gpu.yml --profile full up -d
#
# 2. WITHOUT STREAMLIT (manual UI startup):
# docker compose -f docker-compose.gpu.yml up -d
#
# 3. SINGLE GPU (g5.xlarge - requires stopping LLM for NV-CLIP):
# # Edit CUDA_VISIBLE_DEVICES to share GPU 0
# # Or run services sequentially
#
# =============================================================================
# TROUBLESHOOTING
# =============================================================================
#
# GPU Memory Issues:
# - g5.xlarge (24GB): Can run LLM OR NV-CLIP, not both
# - g5.12xlarge (96GB): Can run all services simultaneously
#
# NIM Authentication:
# - Ensure NGC_API_KEY is set: echo $NGC_API_KEY
# - Get key from: https://ngc.nvidia.com/setup/api-key
#
# Service Logs:
# docker compose -f docker-compose.gpu.yml logs nim-llm
# docker compose -f docker-compose.gpu.yml logs nim-nvclip
#
# Health Status:
# curl http://localhost:8001/v1/models # LLM
# curl -X POST http://localhost:8002/v1/embeddings \
# -H "Content-Type: application/json" \
# -d '{"input":["test"],"model":"nvidia/nvclip"}' # NV-CLIP
#
# =============================================================================
