ROCm vLLM Deployment Skill

Production-ready automation for deploying vLLM inference services on AMD ROCm GPUs using Docker Compose.

Features

• Environment Auto-Check - Detects and repairs missing dependencies
• Model Parameter Detection - Auto-reads config.json for optimal settings
• VRAM Estimation - Calculates memory requirements before deployment
• Secure Token Handling - Never writes tokens to compose files
• Structured Output - All logs and test results saved per-model
• Deployment Reports - Human-readable summary for each deployment
• Health Verification - Automated health checks and functional tests
• Troubleshooting Guide - Common issues and solutions

Environment Prerequisites

Recommended (for production): Add to

~/.bash_profile

# HuggingFace authentication token (required for gated models)
export HF_TOKEN="hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

# Model cache directory (optional)
export HF_HOME="$HOME/models"

# Apply changes
source ~/.bash_profile

Not required for testing: The skill will proceed without these set:

• HF_TOKEN: Optional — public models work without it; gated models fail at download with clear error
• HF_HOME: Optional — defaults to

  /root/.cache/huggingface/hub

Environment Variable Detection

Priority Order:

1. Explicit parameter (highest) — Provided in task/request (e.g.,

   hf_token: "xxx"

   )
2. Environment variable — Already set in shell or from parent process
3. ~/.bash_profile — Source to load variables
4. Default value (lowest) — HF_HOME defaults to

   /root/.cache/huggingface/hub

HF_TOKEN

HF_HOME

/root/.cache/huggingface/hub

Philosophy: Fail fast for configuration errors, fail at download time for authentication errors.

Helper Scripts

Location:

<skill-dir>/scripts/

check-env.sh

Validate and load environment variables before deployment.

Usage:

# Basic check (HF_TOKEN optional, HF_HOME optional with default)
./scripts/check-env.sh

# Strict mode (HF_HOME required, fails if not set)
./scripts/check-env.sh --strict

# Quiet mode (minimal output, for automation)
./scripts/check-env.sh --quiet

# Test with environment variables
HF_TOKEN="hf_xxx" HF_HOME="/models" ./scripts/check-env.sh

Exit Codes:

Note: This script is optional. You can also directly run

source ~/.bash_profile

generate-report.sh

Generate human-readable deployment report after successful deployment.

Usage:

./scripts/generate-report.sh <model-id> <container-name> <port> <status> [model-load-time] [memory-used]

# Example:
./scripts/generate-report.sh \
  "Qwen-Qwen3-0.6B" \
  "vllm-qwen3-0-6b" \
  "8001" \
  "✅ Success" \
  "3.6" \
  "1.2"

Parameters:

model-id

/

-

container-name

port

status

model-load-time

memory-used

Output:

$HOME/vllm-compose/<model-id>/DEPLOYMENT_REPORT.md

Exit Codes:

Integration: This script is automatically called in Phase 7 of the deployment workflow.

Input Schema

${HF_HOME}

/root/.cache/huggingface/hub

${HF_TOKEN}

Output Structure

All deployment artifacts MUST be saved to:

$HOME/vllm-compose/<model-id-slash-to-dash>/

Convert model ID to directory name by replacing

/

-

• openai/gpt-oss-20b

  →

  $HOME/vllm-compose/openai-gpt-oss-20b/

• Qwen/Qwen3-Coder-Next-FP8

  →

  $HOME/vllm-compose/Qwen-Qwen3-Coder-Next-FP8/

Per-model directory structure:

$HOME/vllm-compose/<model-id>/
├── deployment.log          # Full deployment logs (stdout + stderr)
├── test-results.json       # Functional test results (JSON format)
├── docker-compose.yml      # Generated Docker Compose file
├── .env                    # HF_TOKEN environment (chmod 600, optional)
└── DEPLOYMENT_REPORT.md    # Human-readable deployment summary

File requirements:

• deployment.log

  — Capture ALL container logs during deployment

• test-results.json

  — Save API response from functional test request

• DEPLOYMENT_REPORT.md

  — Generated in Phase 7
• All three files MUST exist before marking deployment as complete

Execution Workflow

Phase 0: Environment Check & Auto-Repair

Step 0.1: Load Environment Variables

# Source ~/.bash_profile to load HF_HOME and HF_TOKEN
source ~/.bash_profile

# If HF_HOME is not defined, it defaults to /root/.cache/huggingface/hub

If HF_HOME is not defined in ~/.bash_profile, it defaults to

/root/.cache/huggingface/hub

Step 0.2: Create Output Directory

• Create:

  $HOME/vllm-compose/<model-id>/

Step 0.3: Initialize Logging

• All output →

  $HOME/vllm-compose/<model-id>/deployment.log

Step 0.4: System Checks

• Detect OS and package manager
• Check Python, pip, huggingface_hub
• Check Docker, docker compose
• Check ROCm tools (rocm-smi/amd-smi)
• Check GPU access (/dev/kfd, /dev/dri)
• Check disk space (20GB minimum)

Phase 1: Model Download

Use HF_HOME from Phase 0 (environment variable or default):

# Download model to HF_HOME
huggingface-cli download <model_id> --local-dir "$HF_HOME/hub/models--<org>--<model>"

# Or use snapshot_download via Python:
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='<model_id>', cache_dir='$HF_HOME')"

Authentication Handling:

On Authentication Failure:

echo "ERROR: Model download failed - authentication required"
echo "This model requires a valid HF_TOKEN."
echo ""
echo "Please add to ~/.bash_profile:"
echo "  export HF_TOKEN=\"hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\""
echo "Then run: source ~/.bash_profile"
exit 1

• Locate model path in HF cache:

  $HF_HOME/hub/models--<org>--<model-name>/

• Log download progress to

  deployment.log

Phase 2: Model Parameter Detection

• Read config.json from model
• Auto-detect: max_model_len, hidden_size, num_attention_heads, num_hidden_layers, vocab_size, dtype
• Validate TP size divides attention heads
• Estimate VRAM requirement

Phase 3: Docker Compose Configuration

Generate files in output directory:

• docker-compose.yml →

  $HOME/vllm-compose/<model-id>/docker-compose.yml

  • Mount HF_HOME as volume (read-only for models)
  • NO hardcoded tokens in compose file

• .env →

  $HOME/vllm-compose/<model-id>/.env

  (optional)
  • Contains:

    HF_TOKEN=<value>

  • Permissions:

    chmod 600

  • Only created if user explicitly requests persistent token storage

Volume mount example:

volumes:
  - ${HF_HOME}:/root/.cache/huggingface/hub:ro
  - /dev/kfd:/dev/kfd
  - /dev/dri:/dev/dri

Important: Docker Compose reads

${HF_HOME}

source ~/.bash_profile

Phase 4: Container Launch

Important: Before deploying, pull the latest image to ensure updates:

docker pull rocm/vllm-dev:nightly

Note: Default port is 9999. Before running docker compose, check if port is available:

ss -tlnp | grep :<port>

• Pass HF_TOKEN at runtime: HF_TOKEN=$HF_TOKEN docker compose up -d
• Wait for container initialization

Phase 5: Health Verification

• Check container status
• Test /health endpoint
• Test /v1/models endpoint

Phase 6: Functional Testing

• Run completion test via

  /v1/chat/completions

  API
• Save response to:

  $HOME/vllm-compose/<model-id>/test-results.json

• Verify response contains valid completion
• Log deployment complete → Append to

  deployment.log

• Deployment is complete only when both files exist:

  • deployment.log

  • test-results.json

Phase 7: Deployment Report

Generate human-readable deployment report using the helper script.

Step 7.1: Extract Deployment Metrics

# Parse deployment.log for metrics
MODEL_LOAD_TIME=$(grep -o "model loading took [0-9.]* seconds" deployment.log | grep -o '[0-9.]*' || echo "N/A")
MEMORY_USED=$(grep -o "took [0-9.]* GiB memory" deployment.log | grep -o '[0-9.]*' || echo "N/A")

Step 7.2: Generate Report

# Execute the report generation script
<skill-dir>/scripts/generate-report.sh \
  "<model-id>" \
  "<container-name>" \
  "<port>" \
  "<status>" \
  "$MODEL_LOAD_TIME" \
  "$MEMORY_USED"

# Example:
./scripts/generate-report.sh \
  "Qwen-Qwen3-0.6B" \
  "vllm-qwen3-0-6b" \
  "8001" \
  "✅ Success" \
  "3.6" \
  "1.2"

Output:

$HOME/vllm-compose/<model-id>/DEPLOYMENT_REPORT.md

Report Contents:

• Output structure verification (file checklist)
• Deployment summary table (health, test, metrics)
• Test results (request/response preview)
• Environment configuration
• Quick commands for operations

Completion Criteria:

• DEPLOYMENT_REPORT.md

  exists in output directory
• Report contains all required sections
• All file checks show ✅

Security Best Practices

1. Never commit tokens to version control — Add

   .env

   to

   .gitignore

2. Use .env files with chmod 600 — Restrict access to owner only
3. Mask tokens in logs — Show only first 10 chars:

   ${TOKEN:0:10}...

4. Pass tokens at runtime —

   HF_TOKEN=$HF_TOKEN docker compose up -d

5. Store tokens in ~/.bash_profile — For production environments, set

   HF_TOKEN

   in user's shell config
6. Set token for gated models — HF_TOKEN is validated at download time; set in ~/.bash_profile for production

Troubleshooting

Environment Variables

HF_TOKEN not set

export HF_TOKEN="hf_xxx"

~/.bash_profile

source ~/.bash_profile

HF_HOME not set

/root/.cache/huggingface/hub

export HF_HOME="/path"

~/.bash_profile

~/.bash_profile not found

~/.bash_profile

Changes not taking effect

source ~/.bash_profile

HF_TOKEN provided but download still fails

Model Download

Authentication required

HF_TOKEN

~/.bash_profile

Model not found

Download timeout

Deployment

pip install huggingface_hub

docker compose

render

sudo usermod -aG render $USER

port

gpu_memory_utilization

Cleanup

cd $HOME/vllm-compose/<model-id>
docker compose down

Status Check

Check deployment status and logs:

# View deployment directory
ls -la $HOME/vllm-compose/<model-id>/

# View live logs
tail -f $HOME/vllm-compose/<model-id>/deployment.log

# View test results
cat $HOME/vllm-compose/<model-id>/test-results.json

# Check container status
docker ps | grep <model-id>

# Verify environment variables
echo "HF_TOKEN: ${HF_TOKEN:0:10}..."
echo "HF_HOME: $HF_HOME"

Quick Start (Production)

Step 1: Add environment variables to ~/.bash_profile

# Required: HuggingFace token
export HF_TOKEN="hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

# Recommended: Custom model storage path (production)
export HF_HOME="/data/models/huggingface"

# Apply changes
source ~/.bash_profile

Step 2: Verify environment is ready

# Source ~/.bash_profile to load variables
source ~/.bash_profile

# Expected output:
# === Environment Ready ===
# Summary:
#   HF_TOKEN: hf_xxxxxx...
#   HF_HOME:  /data/models/huggingface

Step 3: Run deployment

# The skill will automatically:
# 1. Source ~/.bash_profile to load HF_HOME and HF_TOKEN
# 2. Use HF_TOKEN and HF_HOME from environment (or ~/.bash_profile, or defaults)
# 3. Proceed without token for public models
# 4. Fail at download time with clear error if gated model requires token

Version History