Build production-ready RunPod serverless endpoints with optimized cold start times. Use when creating or modifying RunPod serverless workers for (1) vLLM-based LLM inference, (2) ComfyUI image/video generation, or (3) custom Python inference. Supports both baked models (fastest cold starts) and dynamic loading (shared models). Generates complete projects including Dockerfiles, worker handlers, startup scripts, and configuration optimized for minimal cold start latency.
Resources
4Install
npx skillscat add avivk5498/my-claude-code-skills/runpod-serverless-builder Install via the SkillsCat registry.
RunPod Serverless Builder
Build end-to-end RunPod serverless endpoints optimized for extremely short cold start times.
Capabilities
Create production-ready RunPod serverless workers for:
- vLLM - High-performance LLM inference
- ComfyUI - Image/video generation with workflow support
- Custom Inference - User-provided Python inference code
Loading Strategies:
- Baked Models: Models embedded in Docker image for fastest cold starts (<5s)
- Dynamic Loading: Models loaded from network storage at runtime (shared across workers)
Quick Start
Use the interactive project generator:
python3 scripts/init_project.pyThis generates a complete project with:
- Optimized Dockerfile
- RunPod handler (worker.py)
- Startup scripts (for dynamic loading)
- Configuration files
- Documentation
Project Generation Workflow
Step 1: Run the Generator
Execute the script and answer prompts:
import subprocess
skill_dir = "/path/to/runpod-serverless-builder"
subprocess.run(["python3", f"{skill_dir}/scripts/init_project.py"])The script prompts for:
- Project name - e.g., "my-vllm-worker"
- Workload type - vLLM, ComfyUI, or Custom
- Loading strategy - Baked or Dynamic
- Model configuration - Model name, quantization, etc.
- Output directory - Where to generate files
Step 2: Customize Generated Files
The generator creates a complete project structure:
my-runpod-worker/
├── Dockerfile # Optimized for cold starts
├── worker.py # RunPod handler function
├── startup.sh # Dynamic loading (if applicable)
├── requirements.txt # Python dependencies
├── .dockerignore # Build optimization
├── .env.example # Environment variables
└── README.md # Project documentationReview and customize:
- worker.py: Modify handler logic, add custom processing
- Dockerfile: Add custom dependencies, adjust configurations
- startup.sh: Add custom initialization steps
- requirements.txt: Add additional Python packages
Step 3: Build and Deploy
# Build Docker image
docker build -t my-worker:latest .
# Push to registry
docker push registry/my-worker:latest
# Deploy to RunPod Dashboard
# 1. Create template with image
# 2. Set environment variables
# 3. Create endpointManual Implementation (Without Generator)
If you prefer manual implementation or need to understand the patterns:
vLLM Worker
Baked Model Approach:
- Copy Dockerfile template:
shutil.copy("assets/dockerfiles/vllm_baked.dockerfile", "Dockerfile")- Copy worker template:
shutil.copy("assets/workers/worker_vllm.py", "worker.py")- Build with model:
docker build -t my-vllm:latest \
--build-arg MODEL_NAME="meta-llama/Llama-3.1-8B-Instruct" \
--build-arg BASE_PATH="/models" \
.Dynamic Loading Approach:
- Copy Dockerfile and startup script:
shutil.copy("assets/dockerfiles/vllm_dynamic.dockerfile", "Dockerfile")
shutil.copy("assets/startup_scripts/startup_vllm.sh", "startup.sh")- Set environment variables in RunPod:
MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
HF_TOKEN=hf_your_token
GPU_MEMORY_UTILIZATION=0.95ComfyUI Worker
Baked Model Approach:
- Use ComfyUI baked template:
shutil.copy("assets/dockerfiles/comfyui_baked.dockerfile", "Dockerfile")
shutil.copy("assets/workers/worker_comfyui.py", "worker.py")- Modify Dockerfile to download models:
# Add model downloads
RUN aria2c -x 16 -s 16 https://huggingface.co/... \
-d /ComfyUI/models/checkpointsDynamic Loading Approach:
- Use dynamic template with startup script:
shutil.copy("assets/dockerfiles/comfyui_dynamic.dockerfile", "Dockerfile")
shutil.copy("assets/startup_scripts/startup_comfyui.sh", "startup.sh")
shutil.copy("assets/config/extra_model_paths.yaml", "extra_model_paths.yaml")Configure network storage paths in extra_model_paths.yaml
Set environment variables:
GITHUB_PAT=ghp_token # For private repos
CUSTOM_NODES=https://github.com/org/node1.git,https://github.com/org/node2.gitCustom Inference Worker
- Use custom templates:
shutil.copy("assets/dockerfiles/custom_inference.dockerfile", "Dockerfile")
shutil.copy("assets/workers/worker_custom.py", "worker.py")- Implement your inference logic in worker.py:
def initialize_model():
# Load your model
return your_model
def handler(job):
model = initialize_model()
# Your inference logic
result = model.predict(job["input"])
return {"result": result}Handler Patterns
vLLM Handler
from vllm import LLM, SamplingParams
llm = None
def initialize_model():
global llm
if llm is None:
llm = LLM(model=MODEL_NAME, gpu_memory_utilization=0.95)
return llm
def handler(job):
model = initialize_model()
messages = job["input"]["messages"]
# Apply chat template
tokenizer = model.get_tokenizer()
prompt = tokenizer.apply_chat_template(messages, tokenize=False)
# Generate
outputs = model.generate([prompt], SamplingParams(...))
return {"text": outputs[0].outputs[0].text}ComfyUI Handler
def update_workflow(workflow, parameters):
# Update prompt node
workflow[parameters["prompt_node_id"]]["inputs"]["text"] = parameters["prompt"]
# Update seed node
workflow[parameters["seed_node_id"]]["inputs"]["seed"] = parameters.get("seed", 42)
return workflow
def handler(job):
# Load workflow JSON
with open(job["input"]["workflow_path"]) as f:
workflow = json.load(f)
# Update with parameters
workflow = update_workflow(workflow, job["input"])
# Execute with ComfyUI API
output = execute_comfyui_workflow(workflow)
return {"image_base64": output}Cold Start Optimization
Key strategies (see references/cold_start_optimization.md for details):
Baked Models Strategy
- Models embedded in image
- Target: <5 second cold starts
- Best for: Small-medium models, latency-critical workloads
Dynamic Loading Strategy
- Models on network storage
- Target: <60 second cold starts
- Best for: Large models, shared across workers
Dockerfile Optimization
# Use BuildKit cache mounts
RUN --mount=type=cache,target=/root/.cache/pip pip install ...
# Order from least to most frequently changing
COPY requirements.txt /
RUN pip install -r requirements.txt
COPY worker.py /
# Combine commands to reduce layers
RUN apt-get update && apt-get install -y pkg1 pkg2 && apt-get cleanWorker Optimization
# Module-level initialization (runs once per container)
MODEL = load_model() # Cached across warm starts
def handler(job):
# MODEL already loaded for warm starts
return MODEL.predict(job["input"])Reference Documentation
Consult reference files for detailed guidance:
references/cold_start_optimization.md- Comprehensive cold start optimization strategiesreferences/vllm_guide.md- vLLM configuration, API patterns, troubleshootingreferences/comfyui_guide.md- ComfyUI workflow management, custom nodes, video workflows
Load references when needed:
# For cold start optimization questions
with open("references/cold_start_optimization.md") as f:
cold_start_guide = f.read()
# For vLLM-specific configuration
with open("references/vllm_guide.md") as f:
vllm_guide = f.read()
# For ComfyUI workflow patterns
with open("references/comfyui_guide.md") as f:
comfyui_guide = f.read()Common Scenarios
Scenario 1: vLLM with Baked Model
User request: "Create a RunPod endpoint for Llama 3.1 8B with the fastest possible cold starts"
Implementation:
- Run init_project.py or copy vllm_baked templates
- Set MODEL_NAME="meta-llama/Llama-3.1-8B-Instruct" in Dockerfile
- Build with model baked in
- Deploy to RunPod
Scenario 2: ComfyUI with Dynamic Loading
User request: "Build a ComfyUI video generation endpoint that loads models from network storage"
Implementation:
- Run init_project.py selecting ComfyUI + Dynamic
- Configure extra_model_paths.yaml for network storage
- Implement workflow update logic in worker.py
- Deploy with CUSTOM_NODES environment variable
Scenario 3: Custom Inference with User Code
User request: "I have a custom object detection model, help me deploy it to RunPod"
Implementation:
- Run init_project.py selecting Custom
- Copy user's model code to inference/ directory
- Implement initialize_model() and handler() in worker.py
- Add dependencies to requirements.txt
- Build and deploy
Troubleshooting
Slow Cold Starts
- Check if models are baked vs downloaded at runtime
- Review Dockerfile layer caching
- Minimize dependencies in requirements.txt
- Consult
references/cold_start_optimization.md
Worker Errors
- Check logs in RunPod dashboard
- Test worker.py locally:
python3 worker.py - Verify environment variables in .env.example
- Check model loading in initialize_model()
Build Failures
- Verify base image compatibility
- Check requirements.txt for conflicting versions
- Test Dockerfile locally:
docker build .
Best Practices
- Always use the generator first - It implements proven patterns
- Start with baked models - Optimize for cold starts, then consider dynamic loading if needed
- Pin dependency versions - Avoid "latest" tags and unpinned packages
- Profile cold starts - Measure and optimize based on actual metrics
- Test locally before deploying - Run worker.py and docker build locally
- Consult references - Load reference docs for detailed guidance on specific topics