Multi-node inference load balancing

This guide describes how to set up a scalable, high-performance multi-node LLM inference cluster using AMD GPUs, supporting efficient horizontal scaling and highly available deployments.

Architecture overview

This solution implements a distributed LLM inference system with three main components:

• Inference pool: Multiple inference nodes running vLLM or SGLang servers on AMD GPUs using tensor parallelism.

• API gateway layer: A unified entry point that distributes requests across the inference pool. This guide demonstrates two options:

  • A LiteLLM-based load balancer - optimized for LLM workloads with built-in observability.

  • An Nginx-based load balancer - a production-grade reverse proxy with high performance.

• Monitoring layer: Prometheus and Grafana for comprehensive metrics collection and visualization, with additional load testing tools.

This architecture allows horizontal scaling by adding more inference nodes while maintaining a single API endpoint for client applications. This architecture supports various model sizes:

• Small models: Can run efficiently on a single GPU.

• Medium models: Typically require two or more GPUs with tensor parallelism.

• Large models: Requires multi-node deployments for high availability and performance.

Tensor Parallelism distributes model layers across multiple GPUs, allowing inference of models too large to fit in a single GPU’s memory. The --tensor-parallel-size (-tp) parameter determines how many GPUs will share the model weights.

Logical diagram

Prerequisites

• Multiple ROCm-compatible nodes with AMD GPUs.

• Docker and Docker Compose installed on all nodes.

• Network connectivity between nodes.

• Models downloaded to a shared or local storage location.

NUMA configuration

For optimal performance, disable automatic NUMA balancing on each node before starting the inference servers:

# Disable automatic NUMA balancing
sudo sh -c 'echo 0 > /proc/sys/kernel/numa_balancing'

# Verify NUMA balancing is disabled (should return 0)
cat /proc/sys/kernel/numa_balancing

Deployment

This section provides step-by-step instructions for deploying components for multi-node inference load balancing.

Project structure

/llm-cluster/
├── nodes/                 # Inference node files
│   ├── docker-compose.yml
├── gateway/               # API Gateway/Load Balancer files
│   ├── litellm
│   │   ├── config.yaml
│   │   └── docker-compose.yml
│   └── nginx
│       ├── docker-compose.yml
│       └── nginx.conf         
├── monitoring/            # Monitoring stack files
│   ├── docker-compose.yml
│   ├── grafana/
│   │   ├── datasources.yml
│   │   ├── Instinct_Dashboard.json
│   │   └── vLLM_Dashboard.json
│   ├── influxdb/
│   ├── prometheus/
│   │   └── prometheus.yml
│   └── scripts/
│       ├── chat-completions-test.js
│       ├── helpers/
│       │   └── openaiGeneric.js
│       ├── prompt-length-test.js
│       ├── ramp-up-test.js
│       └── stress-test.js

Inference pool setup

Perform these actions on each inference node.

1. Create the directory structure:

   mkdir -p ~/llm-cluster/nodes
   cd ~/llm-cluster/nodes
   

2. Create a .env file in the nodes/ folder the with appropriate configuration for your environment:

   NODE_ID=node1               # Unique identifier for this node
   MODEL_PATH=/path/to/models  # Path to local or shared model storage
   MODEL_NAME=Llama-3.1-8B-Instruct  # Model to deploy
   TP_SIZE=4                   # Tensor parallelism degree (number of GPUs to use)
   GPU_DEVICES=0,1,2,3         # GPU devices to use
   PORT=8000                   # Port to expose the inference API
   SHM_SIZE=32GB               # Shared memory size for container
   

3. Create a docker-compose.yml file for the inference nodes. Two options are provided below for different inference backends.

   vLLM example

   services:
     vllm:
       image: rocm/vllm:instinct_main
       container_name: vllm_${NODE_ID:-node1}
       shm_size: ${SHM_SIZE:-32GB}
       ipc: host
       network_mode: host
       devices:
         - /dev/kfd
         - /dev/dri
       group_add:
         - video
       security_opt:
         - seccomp=unconfined
       volumes:
         - ${MODEL_PATH}:/data/models
       environment:
         - ROCR_VISIBLE_DEVICES=${GPU_DEVICES:-0,1,2,3}
       command: >
         vllm serve /data/models/${MODEL_NAME}
         --dtype float16
         --tensor-parallel-size ${TP_SIZE:-4}
         --port ${PORT:-8000}
       restart: unless-stopped
   

   SGLang example

   services:
     sglang:
       image: lmsysorg/sglang:v0.4.6.post2-rocm630
       container_name: sglang_${NODE_ID:-node1}
       shm_size: ${SHM_SIZE:-32GB}
       ipc: host
       network_mode: host
       devices:
         - /dev/kfd
         - /dev/dri
       group_add:
         - video
       security_opt:
         - seccomp=unconfined
       volumes:
         - ${MODEL_PATH}:/data/models
       environment:
         - ROCR_VISIBLE_DEVICES=${GPU_DEVICES:-0,1,2,3}
         - RCCL_MSCCL_ENABLE=0
         - CK_MOE=1
         - HSA_NO_SCRATCH_RECLAIM=1
       command: >
         python3 -m sglang.launch_server
         --model /data/models/${MODEL_NAME}
         --tp ${TP_SIZE:-4}
         --trust-remote-code
         --port ${PORT:-8000}
         --enable-metrics
       restart: unless-stopped
   

4. Start the inference services:

   docker compose up -d
   

API gateway setup

On the API gateway node, create the gateway directory structure:

mkdir -p ~/llm-cluster/gateway
cd ~/llm-cluster/gateway

Choose one of the following gateway options based on your requirements.

Option 1: LiteLLM-based load balancer

LiteLLM provides specialized routing, load balancing, and observability for LLM API calls, supporting multiple LLM providers and models through a unified OpenAI-compatible interface.

1. Create docker-compose.yml for LiteLLM:

   services:
     litellm:
       image: ghcr.io/berriai/litellm:main-stable
       container_name: litellm_gateway
       network_mode: host
       volumes:
         - ./config.yaml:/app/config.yaml
       command: ["--config", "/app/config.yaml", "--port", "4000", "--num_workers", "8"]
       environment:
         LITELLM_MASTER_KEY: "${LITELLM_MASTER_KEY}"
       env_file: .env
       restart: unless-stopped
       logging:
         driver: "json-file"
         options:
           max-size: "10m"
           max-file: "3"
   

2. Create config.yaml to define the model routing configuration:

   model_list:
   - model_name: DeepSeek-R1
      litellm_params:
         model: openai/deepseek-ai/DeepSeek-R1
         api_base: http://node0:8000/v1
   
   - model_name: DeepSeek-R1
      litellm_params:
         model: openai/deepseek-ai/DeepSeek-R1
         api_base: http://node1:8000/v1
   
   # Add additional nodes as needed
   # - model_name: DeepSeek-R1
   #   litellm_params:
   #     model: openai/deepseek-ai/DeepSeek-R1
   #     api_base: http://nodeN:8000/v1
   
   # Configure load balancing
   router_settings:
   routing_strategy: least-busy  # Distributes requests to least busy nodes
   num_retries: 3                # Number of retries if a request fails
   timeout: 300                  # Request timeout in seconds
   

3. Create .env file with your API key:

   LITELLM_MASTER_KEY=sk-1234
   

   Note

   For production environments, replace the default key with a strong, randomized value.

4. Start the LiteLLM gateway:

   docker compose up -d
   

5. Verify that all LLM endpoints are healthy:

   curl -X 'GET' \
   'http://localhost:4000/health' \
   -H 'accept: application/json' \
   -H 'Authorization: Bearer sk-1234' | jq
   

   Expected output

   {
   "healthy_endpoints": [
      {
         "model": "openai/deepseek-ai/DeepSeek-R1",
         "api_base": "http://node0:8000/v1"
      },
      {
         "model": "openai/deepseek-ai/DeepSeek-R1",
         "api_base": "http://node1:8000/v1"
      },
      {
         "model": "openai/deepseek-ai/DeepSeek-R1",
         "api_base": "http://node2:8000/v1"
      },
      {
         "model": "openai/deepseek-ai/DeepSeek-R1",
         "api_base": "http://node3:8000/v1"
      }
   ],
   "unhealthy_endpoints": [],
   "healthy_count": 4,
   "unhealthy_count": 0
   }
   

LiteLLM monitoring options

LiteLLM provides several monitoring and observability options:

• Basic logging: Available in the open source version, provides request/response logging and basic metrics

• Callback integrations: LiteLLM supports custom callbacks for advanced monitoring with tools like:

  • LangFuse

  • Helicone

  • LangSmith

  • Custom callback handlers

This guide uses the open source version of LiteLLM with an internal Prometheus/Grafana stack for system-level monitoring. If you need LLM-specific tracing and observability, consider exploring the callback integrations.

Option 2: Nginx-based load balancer

Nginx provides a high-performance, scalable HTTP server and reverse proxy that can efficiently distribute traffic across multiple inference nodes.

1. Create nginx.conf with the following configuration:

   worker_processes auto;
   worker_rlimit_nofile 65535;
   events {
      worker_connections 65535;
   }
   
   http {
      include       mime.types;
      default_type  application/octet-stream;
      sendfile      on;
      keepalive_timeout 65;
   
      # Define upstream server group
      upstream vllm_pool {
         # Use least_conn for distributing traffic based on least number of current connections
         least_conn;
         
         # Add inference server entries - update with your node hostnames/IPs
         server node0:8000;
         server node1:8000;
         # Add additional nodes as needed
         # server nodeN:8000;
         
         keepalive 32;
      }
   
      server {
         listen 80;
         
         # Health check endpoint
         location /health {
               return 200 'healthy\n';
               add_header Content-Type text/plain;
         }
   
         # API endpoint for frontend clients
         location / {
               proxy_pass http://vllm_pool;
               proxy_http_version 1.1;
               proxy_set_header Connection "";
               proxy_set_header Host $host;
               proxy_set_header X-Real-IP $remote_addr;
               proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
               
               # Timeouts for long-running inference requests
               proxy_connect_timeout 300s;
               proxy_read_timeout 300s;
               proxy_send_timeout 300s;
               
               # Buffer settings for large responses
               proxy_buffer_size 16k;
               proxy_buffers 8 16k;
               proxy_busy_buffers_size 32k;
         }
      }
   }
   

2. Create docker-compose.yml for Nginx:

   services:
   nginx:
      image: nginx:latest
      container_name: nginx_gateway
      network_mode: host
      volumes:
         - ./nginx.conf:/etc/nginx/nginx.conf:ro
      restart: unless-stopped
      logging:
         driver: "json-file"
         options:
         max-size: "10m"
         max-file: "3"
   

3. Start the Nginx gateway:

   docker compose up -d
   

Monitoring Nginx gateway

To enable monitoring for your Nginx gateway, add the nginx-prometheus-exporter:

1. Update docker-compose.yml to include the exporter:

   services:
   nginx:
      # ...existing nginx configuration...
   
   nginx-exporter:
      image: nginx/nginx-prometheus-exporter:latest
      container_name: nginx_exporter
      command:
         - --nginx.scrape-uri=http://localhost/stub_status
      network_mode: host
      restart: unless-stopped
      depends_on:
         - nginx
   

2. Add a status endpoint to nginx.conf inside the server block:

   location /metrics {
      stub_status on;
      access_log off;
      allow 127.0.0.1;
      deny all;
   }
   

Monitoring stack setup

Perform these steps on the monitoring node.

1. Create the monitoring directory structure:

   mkdir -p ~/llm-cluster/monitoring/{prometheus,grafana,influxdb}
   cd ~/llm-cluster/monitoring
   

2. Set appropriate permissions for Grafana and InfluxDB data directories:

   # Set permissions to allow container processes to write data
   chmod 777 ~/llm-cluster/monitoring/grafana
   chmod 777 ~/llm-cluster/monitoring/influxdb
   

3. Create docker-compose.yml for the monitoring stack:

   services:
   # Check https://hub.docker.com/r/rocm/device-metrics-exporter/tags for the latest version
   device-metrics-exporter:
      image: rocm/device-metrics-exporter:v1.3.0-beta.1
      container_name: device-metrics-exporter
      restart: unless-stopped
      group_add:
         - video    
      volumes:
         - ./config.json:/etc/metrics/config.json
      devices:
         - /dev/kfd
         - /dev/dri
      ports:
         - "5000:5000"
   
   prometheus:
      image: prom/prometheus:latest
      container_name: prometheus
      volumes:
         - ./prometheus/prometheus.yml:/etc/prometheus/prometheus.yml
      command:
         - '--config.file=/etc/prometheus/prometheus.yml'
      ports:
         - "9090:9090"
      restart: unless-stopped
   
   influxdb:
      image: influxdb:1.11.8
      container_name: influxdb
      ports:
         - "8086:8086"
      environment:
         - INFLUXDB_DB=k6
         - INFLUXDB_ADMIN_USER=admin
         - INFLUXDB_ADMIN_PASSWORD=admin
      volumes:
         - ./influxdb:/var/lib/influxdb
   
   grafana:
      image: grafana/grafana:latest
      container_name: grafana
      volumes:
         - ./grafana/datasources.yml:/etc/grafana/provisioning/datasources/datasources.yml
         - ./grafana:/var/lib/grafana
      environment:
         - GF_SECURITY_ADMIN_PASSWORD=${GRAFANA_ADMIN_PASSWORD:-admin}
      ports:
         - "3000:3000"
      depends_on:
         - prometheus
      restart: unless-stopped
   

4. Create prometheus/prometheus.yml to configure metrics collection:

   global:
   scrape_interval: 15s
   
   scrape_configs:
   # Host OS metrics
   - job_name: 'node'
      static_configs:
      - targets: ['localhost:9100']
   
   # Inference servers
   - job_name: 'vllm'
      metrics_path: /metrics
      scrape_interval: 15s
      static_configs:
         - targets: ['node0:8000', 'node1:8000'] # Add additional nodes as needed
         labels:
            service: 'vllm'
   
   # Nginx Gateway metrics (if using Nginx with nginx-prometheus-exporter)
   - job_name: 'nginx'
      scrape_interval: 15s
      metrics_path: /metrics
      static_configs:
         - targets: ['localhost:9113']
      relabel_configs:
         - source_labels: [__address__]
         target_label: instance
         replacement: 'nginx-gateway'
   
   # AMD GPU device metrics
   - job_name: 'amd_gpu_metrics'
      scrape_interval: 5s
      metrics_path: /metrics
      static_configs:
         - targets: ['node0:5000', 'node1:5000']
         labels:
            service: 'amd_gpu_metrics'        
   

   Note

   Replace node0 and node1 with the actual hostnames or IP addresses of your inference nodes. When running Prometheus in a docker container, change instances of localhost to host.docker.internal.

5. Create grafana/datasources.yml to configure the Prometheus data source:

   apiVersion: 1
   
   datasources:
   - name: Prometheus
      type: prometheus
      access: proxy
      url: http://prometheus:9091
      isDefault: true
   
   - name: InfluxDB
      type: influxdb
      access: proxy
      url: http://influxdb:8086
      database: k6
      user: admin
      password: admin
      editable: true    
   

6. Start the monitoring services:

   docker compose up -d
   

Testing and performance evaluation

Once your multi-node inference system is deployed, you can validate its functionality and evaluate its performance.

Testing with LiteLLM gateway

Send a test request to the LiteLLM endpoint:

curl http://localhost:4000/v1/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-1234" \
  -d '{"model": "DeepSeek-R1", "prompt": "What is AMD Instinct?", "max_tokens": 256, "temperature": 0.0}'

Testing with Nginx gateway

Send a test request to the Nginx endpoint:

curl http://localhost/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "DeepSeek-R1", "prompt": "What is AMD Instinct?", "max_tokens": 256, "temperature": 0.0}'

Expected output format (content may vary):

{
  "text": [
    "What is AMD Instinct? AMD Instinct is a line of high-performance computing (HPC) and 
    artificial intelligence (AI) accelerators designed for datacenter and cloud computing 
    applications. It is based on AMDs Radeon Instinct architecture, which is optimized for HPC
    and AI workloads. AMD Instinct accelerators are designed to provide high-performance 
    computing and AI acceleration for a wide range of applications, including scientific simulations, 
    data analytics, machine learning, and deep learning.
     
    AMD Instinct accelerators are based on AMDs Radeon Instinct architecture, which is designed 
    to provide high-performance computing and AI acceleration. They are built on a 7nm process node 
    and feature a high-performance GPU core, as well as a large amount of memory and bandwidth to 
    support high-performance computing and AI workloads.
     
    AMD Instinct accelerators are designed to be used in a variety of applications, including:
    Scientific simulations: AMD Instinct accelerators can be used to accelerate complex scientific 
    simulations, such as weather forecasting, fluid dynamics, and molecular dynamics.
    Data analytics: AMD Instinct accelerators can be used to accelerate data analytics workloads,
    such as data compression, data encryption, and data mining.
    Machine learning: AMD Instinct accelerators can be used to accelerate machine learning workloads"
  ]
}

Performance testing with Apache Bench

Apache Bench (ab) is a lightweight tool for benchmarking HTTP servers, ideal for quick performance evaluation.

Installation options

Option 1: Install Apache Bench Locally

sudo apt-get update
sudo apt-get install apache2-utils

Option 2: Run Apache Bench in a Container

docker run -it --rm \
  --shm-size=8GB \
  --ipc=host \
  --network=host \
  --entrypoint bash \
  ubuntu/apache2:2.4-22.04_beta

Running Apache Bench tests

1. Create a request payload file:

   cat > postdata << EOF
   {"model": "DeepSeek-R1", "prompt": "What is AMD Instinct?", "max_tokens": 256, "temperature": 0.0}
   EOF
   

2. Run the benchmark with desired concurrency and request count:

   ab -n 1000 -c 100 -T application/json -p postdata -H "Authorization: Bearer sk-1234" http://localhost:4000/v1/completions
   

Key parameters:

• -n 1000: Total number of requests to perform

• -c 100: Number of concurrent requests

• -T application/json: Content-type header for POST data

• -p postdata: File containing data to POST

• -H: Additional header for authentication

Sample performance test commands

Here are examples of commands to test different models and configurations:

# Test with Llama-3.1-8B-Instruct
ab -n 20000 -c 2000 -T application/json -p postdata http://localhost:80/v1/completions

# Test with Llama-3.1-405B-Instruct
ab -n 20000 -c 2000 -T application/json -p postdata http://localhost:80/v1/completions

# Test with DeepSeek-R1
ab -n 20000 -c 2000 -T application/json -p postdata http://localhost:80/v1/completions

Advanced load testing with k6

For more sophisticated load testing scenarios, Grafana k6 offers enhanced capabilities including detailed metrics collection and realistic user simulation. The test scripts used in this section are available to download from ROCm/gpu-cluster-networking

Installing k6

Option 1: Install k6 locally

apt install -y k6

For additional installation options, refer to the official k6 installation guide.

Option 2: Run k6 in a Container

docker run --rm -i \
  --network=host \
  -v ${PWD}/scripts:/scripts \
  -e "OPENAI_URL=http://localhost:4000" \
  -e "API_KEY=sk-1234" \
  -e "MODEL_NAME=DeepSeek-R1" \
  grafana/k6 run /scripts/chat-completions-test.js

Setting up k6

Configure environment variables for the test scripts:

cd ~/llm-cluster/monitoring/scripts
cat > .env << EOL
export OPENAI_URL=http://localhost:4000  # Use your LiteLLM or Nginx endpoint
export API_KEY=sk-1234                   # API key if required by your gateway
export MODEL_NAME=DeepSeek-R1            # Your deployed model name
EOL

source .env

Running k6 test scripts

The repository includes several specialized test scripts for different testing scenarios:

Chat completions test

k6 run --out influxdb=http://localhost:8086/k6 chat-completions-test.js

Ramp-up test

k6 run --out influxdb=http://localhost:8086/k6 ramp-up-test.js

Stress test

k6 run --out influxdb=http://localhost:8086/k6 stress-test.js

Prompt length test

k6 run --out influxdb=http://localhost:8086/k6 prompt-length-test.js

On completion, k6 will provide a summary similar to this:

$ k6 run --out influxdb=http://localhost:8086 scripts/chat-completions-test.js

         /\      Grafana   /‾‾/
    /\  /  \     |\  __   /  /
   /  \/    \    | |/ /  /   ‾‾\
  /          \   |   (  |  (‾)  |
 / __________ \  |_|\_\  \_____/

     execution: local
        script: scripts/chat-completions-test.js
        output: InfluxDBv1 (http://localhost:8086)

     scenarios: (100.00%) 1 scenario, 5 max VUs, 1m30s max duration (incl. graceful stop):
              * default: 5 looping VUs for 1m0s (gracefulStop: 30s)

  █ THRESHOLDS

    http_req_duration
    ✓ 'p(95)<5000' p(95)=1.66s

    http_req_failed
    ✓ 'rate<0.01' rate=0.00%


  █ TOTAL RESULTS

    checks_total.......................: 170     2.64314/s
    checks_succeeded...................: 100.00% 170 out of 170
    checks_failed......................: 0.00%   0 out of 170

    ✓ is status 200
    ✓ has valid JSON response

    CUSTOM
    completion_tokens...................: avg=100       min=100      med=100       max=100       p(90)=100       p(95)=100
    prompt_tokens.......................: avg=26        min=26       med=26        max=26        p(90)=26        p(95)=26
    tokens_per_second...................: avg=83.173393 min=57.87037 med=87.565674 max=91.324201 p(90)=90.546921 p(95)=90.810037
    total_tokens........................: avg=126       min=126      med=126       max=126       p(90)=126       p(95)=126

    HTTP
    http_req_duration...................: avg=1.21s     min=1.09s    med=1.14s     max=1.72s     p(90)=1.44s     p(95)=1.66s
      { expected_response:true }........: avg=1.21s     min=1.09s    med=1.14s     max=1.72s     p(90)=1.44s     p(95)=1.66s
    http_req_failed.....................: 0.00% 0 out of 85
    http_reqs...........................: 85    1.32157/s

    EXECUTION
    iteration_duration..................: avg=3.67s     min=2.16s    med=3.62s     max=5.35s     p(90)=4.82s     p(95)=4.96s
    iterations..........................: 85    1.32157/s
    vus.................................: 1     min=1       max=5
    vus_max.............................: 5     min=5       max=5

    NETWORK
    data_received........................: 87 kB 1.3 kB/s
    data_sent............................: 31 kB 477 B/s

running (1m04.3s), 0/5 VUs, 85 complete and 0 interrupted iterations
default ✓ [======================================] 5 VUs  1m0s

Viewing k6 test results

After running the tests, you can view the results in Grafana:

1. Open Grafana at http://<your-monitoring-node-ip>:3000

2. Log in with your credentials (default: admin/admin, unless changed via GRAFANA_ADMIN_PASSWORD environment variable)

3. Access the k6 dashboard by importing the dashboard ID 14801 or by navigating to the pre-configured dashboard if available. The dashboard can be found at: https://grafana.com/grafana/dashboards/14801-k6-dashboard/

The k6 dashboard provides detailed metrics about request rates, response times, errors, and other performance indicators that help you understand your system’s behavior under load.

Monitoring and visualization

Available dashboards

The monitoring stack includes pre-configured Grafana dashboards for comprehensive system visibility. These dashboards are provided in the repository’s examples/llm-cluster/monitoring/grafana directory:

AMD Instinct dashboard (Instinct_Dashboard.json): Monitors GPU performance metrics including temperature, utilization, memory usage, and power consumption. Also available at AMD Instinct Single Node Dashboard.

vLLM dashboard (vLLM_Dashboard.json): Provides insights into vLLM server performance, including request throughput, latency metrics, and queue statistics.

Additional recommended dashboards for comprehensive monitoring:

• k6 dashboard: Visualizes load test results with detailed performance metrics. Available for import into Grafana with ID 14801 or at https://grafana.com/grafana/dashboards/14801-k6-dashboard/.

• vLLM reference dashboard: Official dashboard from the vLLM project for detailed inference metrics. Available at vllm-project/vllm.

• NGINX dashboard: Official dashboard for the NGINX Prometheus exporter. https://grafana.com/grafana/dashboards/12767-nginx/

For instructions on importing dashboards into your Grafana instance, follow the official Grafana Dashboard Import Guide.

Performance optimization recommendations

Achieving optimal performance for your multi-node inference deployment requires experimentation and continuous monitoring. This section provides recommendations for tuning your setup based on your specific workload characteristics.

Compare different configurations

To identify the optimal setup for your specific use case, systematically test different configurations.

• Load balancer options

  • LiteLLM: Generally provides better handling of LLM-specific requirements like streaming responses and specialized routing

  • Nginx: Often delivers higher raw throughput for simple completion requests and offers more configuration flexibility

• Inference servers

  • vLLM: https://docs.vllm.ai/

  • SGLang: https://docs.sglang.ai/

  • TGI: https://huggingface.co/docs/text-generation-inference/index

• Inference configuration

  • Test different tensor parallel sizes to find the optimal balance between throughput and latency.

  • Experiment with batch sizes (--max-batch-size in vLLM) to increase throughput for concurrent requests.

  • Try different quantization options to improve memory efficiency.

Using historical performance data

You can use the monitoring setup in this guide to review stored historical performance data and track changes over time:

• Establish performance baselines

  • Run benchmark tests after initial setup to establish baseline performance metrics.

  • Document key metrics like tokens per second, request latency, and GPU utilization.

• Track performance trends

  • Set up Grafana dashboards with time series views of key metrics.

  • Create alerts for significant deviations from established baselines.

System-level optimizations

Beyond the application components themselves, consider these system-level optimizations:

• Network configuration

  • Ensure nodes have sufficient network bandwidth for model weight synchronization.

  • Consider using dedicated network interfaces for inter-node communication.

• Host OS tuning

  • Adjust kernel parameters related to networking and memory management.

  • The NUMA configuration mentioned earlier in this guide is just one example.

You can find more information on system optimization at these links:

• System optimization guides: https://rocm.docs.amd.com/en/latest/how-to/system-optimization/index.html

• Performance guides: https://rocm.docs.amd.com/en/latest/how-to/gpu-performance/mi300x.html

• Instinct single node networking: https://instinct.docs.amd.com/projects/gpu-cluster-networking/en/latest/how-to/single-node-config.html

• Instinct multi-node networking: https://instinct.docs.amd.com/projects/gpu-cluster-networking/en/latest/how-to/multi-node-config.html

Cost-performance balance

When scaling your cluster, consider both performance and resource utilization:

• Right-sizing

  • Use Grafana dashboards to identify under-utilized resources.

  • Scale the number of nodes based on actual usage patterns and SLAs.

• Workload scheduling

  • Consider dedicating specific nodes to different models based on usage patterns.

  • Use metrics to identify peak usage times and scale accordingly.

By systematically testing configurations and leveraging the monitoring data, you can continuously optimize your multi-node inference setup to achieve the best balance of performance, reliability, and resource efficiency.

Repository resources

All configuration files, scripts, and dashboards referenced in this guide are available in the ROCm GPU Cluster Networking GitHub repository:

ROCm/gpu-cluster-networking

The repository includes:

• Docker Compose files for inference nodes (vLLM and SGLang examples)

• API Gateway configurations (LiteLLM and Nginx examples)

• Monitoring stack setup with Prometheus, Grafana, and InfluxDB

• Grafana dashboards for AMD Instinct GPUs and vLLM

• Benchmark scripts for Apache Bench and k6

• Example configuration files and setup scripts