vllm-project.github.io/_posts/2025-04-11-transformers-backend.md

layout	title	author	image	thumbnail-img	share-img
post	Transformers backend integration in vLLM	The Hugging Face Team	/assets/figures/transformers-backend/transformers-backend.png	/assets/figures/transformers-backend/transformers-backend.png	/assets/figures/transformers-backend/transformers-backend.png

The Hugging Face Transformers library offers a flexible, unified interface to a vast ecosystem of model architectures. From research to fine-tuning on custom dataset, transformers is the go-to toolkit for all.

But when it comes to deploying these models at scale, inference speed and efficiency often take center stage. Enter vLLM, a library engineered for high-throughput inference, pulling models from the Hugging Face Hub and optimizing them for production-ready performance.

A recent addition to the vLLM codebase enables leveraging transformers as a backend to run models. vLLM will therefore optimize throughput/latency on top of existing transformers architectures. In this post, we’ll explore how vLLM leverages the transformers backend to combine flexibility with efficiency, enabling you to deploy state-of-the-art models faster and smarter.

Transformers and vLLM: Inference in Action

Let’s start with a simple text generation task using the meta-llama/Llama-3.2-1B model to see how these libraries stack up.

Infer with transformers

The transformers library shines in its simplicity and versatility. Using its pipeline API, inference is a breeze:

from transformers import pipeline

pipe = pipeline("text-generation", model="meta-llama/Llama-3.2-1B")
result = pipe("The future of AI is")

print(result[0]["generated_text"])

This approach is perfect for prototyping or small-scale tasks, but it’s not optimized for high-volume inference or low-latency deployment.

Infer with vLLM

vLLM takes a different track, prioritizing efficiency with features like PagedAttention (a memory-efficient attention mechanism) and dynamic batching. Here’s the same task in vLLM:

from vllm import LLM, SamplingParams

llm = LLM(model="meta-llama/Llama-3.2-1B")
params = SamplingParams(max_tokens=20)
outputs = llm.generate("The future of AI is", sampling_params=params)
print(f"Generated text: {outputs[0].outputs[0].text}")

vLLM’s inference is noticeably faster and more resource-efficient, especially under load. For example, it can handle thousands of requests per second with lower GPU memory usage.

vLLM’s Deployment Superpower: OpenAI Compatibility

Beyond raw performance, vLLM offers an OpenAI-compatible API, making it a drop-in replacement for external services. Launch a server:

vllm serve meta-llama/Llama-3.2-1B

Then query it with curl:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "meta-llama/Llama-3.2-1B", "prompt": "San Francisco is a", "max_tokens": 7, "temperature": 0}'

Or use Python’s OpenAI client:

from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")
completion = client.completions.create(
    model="meta-llama/Llama-3.2-1B",
    prompt="San Francisco is a",
    max_tokens=7,
    temperature=0
)
print("Completion result:", completion.choices[0].text)

This compatibility slashes costs and boosts control, letting you scale inference locally with vLLM’s optimizations.

Why do we need the transformers backend?

The transformers library is optimized for contributions and addition of new models. Adding a new model to vLLM on the other hand is a little more involved.

In the ideal world, we would be able to use the new model in vLLM as soon as it is added to transformers. With the integration of the transformers backend, we step towards that ideal world.

Here is the official documentation on how to make your transformers model compatible with vLLM for the integration to kick in. We followed this and made modeling_gpt2.py compatible with the integration! You can follow the changes in this transformers pull request.

For a model already in transformers (and compatible with vLLM), this is what we would need to:

llm = LLM(model="new-transformers-model", model_impl="transformers")

[!NOTE] It is not a strict necessity to add model_impl parameter. vLLM switches to the transformers implementation on its own if the model is not natively supported in vLLM.

Or for a custom model from the Hugging Face Hub:

llm = LLM(model="custom-hub-model", model_impl="transformers", trust_remote_code=True)

This backend acts as a bridge, marrying transformers’ plug-and-play flexibility with vLLM’s inference prowess. You get the best of both worlds: rapid prototyping with transformers and optimized deployment with vLLM.

Case Study: Helium

Kyutai Team’s Helium is not yet supported by vLLM. You might want to run optimized inference on the model with vLLM, and this is where the transformers backend shines.

Let’s see this in action:

vllm serve kyutai/helium-1-preview-2b --model-impl transformers

Query it with the OpenAI API:

from openai import OpenAI

openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

completion = client.completions.create(model="kyutai/helium-1-preview-2b", prompt="What is AI?")
print("Completion result:", completion)

Here, vLLM efficiently processes inputs, leveraging the transformers backend to load kyutai/helium-1-preview-2b seamlessly. Compared to running this natively in transformers, vLLM delivers lower latency and better resource utilization.

By pairing Transformers’ model ecosystem with vLLM’s inference optimizations, you unlock a workflow that’s both flexible and scalable. Whether you’re prototyping a new model, deploying a custom creation, or scaling a multimodal app, this combination accelerates your path from research to production.