PyLet

A simple distributed task execution system for GPU servers. Like Ray/K8s but much simpler.

Install

pip install pylet

For development:

git clone https://github.com/ServerlessLLM/pylet.git
cd pylet
pip install -e ".[dev]"

Quick Start

CLI

# Terminal 1: Start head node
pylet start

# Terminal 2: Start worker node with GPUs
pylet start --head localhost:8000 --gpu-units 4

# Terminal 3: Submit an instance
pylet submit 'vllm serve Qwen/Qwen2.5-1.5B-Instruct --port $PORT' \
    --gpu-units 1 --name my-vllm

# Check status
pylet get-instance --name my-vllm

# Get endpoint for inference
pylet get-endpoint --name my-vllm
# Output: 192.168.1.10:15600

# View logs
pylet logs <instance-id>

# Cancel
pylet cancel <instance-id>

Python API

import pylet

# Connect to head node
pylet.init()  # or pylet.init("http://head:8000")

# Submit an instance
instance = pylet.submit(
    "vllm serve Qwen/Qwen2.5-1.5B-Instruct --port $PORT",
    name="my-vllm",
    gpu=1,
    memory=4096,
)

# Wait for it to start
instance.wait_running()
print(f"Endpoint: {instance.endpoint}")

# Get logs
print(instance.logs())

# Cancel when done
instance.cancel()
instance.wait()

For local testing:

import pylet

with pylet.local_cluster(workers=2, gpu_per_worker=1) as cluster:
    instance = pylet.submit("nvidia-smi", gpu=1)
    instance.wait()
    print(instance.logs())

Async API available via import pylet.aio as pylet.

See examples/README.md for more detailed examples including vLLM and SGLang.

Commands

Key Features

• Simple: No containers, no complex configs. Just pylet start and pylet submit.
• GPU-aware: Automatic GPU allocation via CUDA_VISIBLE_DEVICES.
• Fine-grained GPU control: Request specific GPUs by index, enable GPU sharing for daemons.
• Service discovery: Instances get a PORT env var; endpoint available via get-endpoint.
• Real-time logs: Stream logs from running instances.
• Graceful shutdown: SIGTERM with configurable grace period before SIGKILL.

Why PyLet?

PyLet is the greatest common denominator for GPU cluster orchestration - minimal, reliable, and unopinionated. It manages one thing: instances (processes with GPU allocation).

• No pods, replicas, services, or deployments - just instances
• Applications compose instances however they need (via labels, custom scheduling)
• Fine-grained GPU scheduling because research workloads need it
• Simple enough to understand, reliable enough to depend on

See docs/architecture.md for design details. See CONTRIBUTING.md for contribution philosophy.

Requirements

• Python 3.9+
• Linux (tested on Ubuntu)

License

Apache 2.0