flyteplugins-vllm 2.1.3

vLLM plugin for flyte

Union vLLM Plugin

Serve large language models using vLLM with Flyte Apps.

This plugin provides the VLLMAppEnvironment class for deploying and serving LLMs using vLLM.

Installation

pip install --pre flyteplugins-vllm

Usage

import flyte
import flyte.app
from flyteplugins.vllm import VLLMAppEnvironment

# Define the vLLM app environment
vllm_app = VLLMAppEnvironment(
    name="my-llm-app",
    model="s3://your-bucket/models/your-model",
    model_id="your-model-id",
    resources=flyte.Resources(cpu="4", memory="16Gi", gpu="L40s:1"),
    stream_model=True,  # Stream model directly from blob store to GPU
    scaling=flyte.app.Scaling(
        replicas=(0, 1),
        scaledown_after=300,
    ),
)

if __name__ == "__main__":
    flyte.init_from_config()
    app = flyte.serve(vllm_app)
    print(f"Deployed vLLM app: {app.url}")

Features

• Streaming Model Loading: Stream model weights directly from object storage to GPU memory, reducing startup time and disk requirements.
• OpenAI-Compatible API: The deployed app exposes an OpenAI-compatible API for chat completions.
• Auto-scaling: Configure scaling policies to scale up/down based on traffic.
• Tensor Parallelism: Support for distributed inference across multiple GPUs.