Instructions to use Cloudflare/clef-flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Cloudflare/clef-flash with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Cloudflare/clef-flash")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Cloudflare/clef-flash")
model = AutoModelForMultimodalLM.from_pretrained("Cloudflare/clef-flash", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Cloudflare/clef-flash with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Cloudflare/clef-flash"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Cloudflare/clef-flash",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Cloudflare/clef-flash

SGLang

How to use Cloudflare/clef-flash with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Cloudflare/clef-flash" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Cloudflare/clef-flash",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Cloudflare/clef-flash" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Cloudflare/clef-flash",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Cloudflare/clef-flash with Docker Model Runner:
```
docker model run hf.co/Cloudflare/clef-flash
```

Support / Roadmap for serving Clef-Flash via vLLM?

by Vishva007 - opened 2 days ago

Discussion

Vishva007

2 days ago

•

edited 2 days ago

Hi @Cloudflare team,

Thanks for releasing Clef-Flash! The joint schema evaluation approach and single forward-pass probability output design are exceptional for high-throughput decision and classification pipelines.

Since standard deployment in production environments relies heavily on vLLM (for PagedAttention, continuous batching, and high concurrency), I wanted to check on vLLM compatibility:

Custom Architecture Support: Since clef-flash relies on custom code (joint_schema_model.py / systemone) and evaluates typed questions via direct hidden-state readouts rather than token-by-token autoregressive generation, is there any current pathway or plan to register ClefForConditionalGeneration / its joint classification heads as an out-of-the-box model architecture in vLLM?
Forward-Pass / Embedding API: In vLLM, classification and embedding models typically use the pooling/forward-pass endpoints rather than /v1/completions. Has the team explored integrating systemone schema scoring on top of vLLM’s forward/pooling engine?
Current Serving Recommendations: Outside of running native Hugging Face Transformers + FastAPI/Triton, what is your recommended high-throughput serving stack for serving Clef-Flash at scale?

Appreciate any guidance or pointers from the team or community!

ironbcc

1 day ago

@Vishva007 If stock vLLM support is the immediate requirement, there is another implementation approach: score the next-token logits restricted to the decision's option keys, rather than serving a custom joint head. I built Seb-9B around that approach: https://huggingface.co/ironbcc/seb-9b

The card includes a vLLM chat API example with constrained choices and option logprobs. It handles one question per request, with up to 20 choice options; that differs from Clef's joint multi-question design. Released-checkpoint GPU throughput measurements are still pending, so I can't claim it solves your high-concurrency requirement yet. It may be useful as a stock-vLLM baseline while the Clef integration question is worked through.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment