Instructions to use dill-lab/pils-32-llama2-chat-7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dill-lab/pils-32-llama2-chat-7b with Transformers:

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

pipe = pipeline("text-generation", model="dill-lab/pils-32-llama2-chat-7b")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("dill-lab/pils-32-llama2-chat-7b", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dill-lab/pils-32-llama2-chat-7b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dill-lab/pils-32-llama2-chat-7b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dill-lab/pils-32-llama2-chat-7b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/dill-lab/pils-32-llama2-chat-7b

SGLang

How to use dill-lab/pils-32-llama2-chat-7b 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 "dill-lab/pils-32-llama2-chat-7b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dill-lab/pils-32-llama2-chat-7b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "dill-lab/pils-32-llama2-chat-7b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dill-lab/pils-32-llama2-chat-7b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use dill-lab/pils-32-llama2-chat-7b with Docker Model Runner:
```
docker model run hf.co/dill-lab/pils-32-llama2-chat-7b
```

Install dependencies

pip install gradio
pip install git+https://github.com/dill-lab/PILS

Run the demo

import gradio as gr
import torch
from pils.models import InversionFromHiddenStatesModel


MODEL = InversionFromHiddenStatesModel.from_pretrained(
    "murtaza/pils-32-llama2-chat-7b")
MODEL.embedder_no_grad=True
MODEL.embedder.max_new_tokens = 64
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
MODEL = MODEL.to(DEVICE)

def invert(user_prompt):
    global inp
    sys_prompt = ''
    inp = MODEL.embedder_tokenizer.apply_chat_template(conversation=[
        {"role": "system", "content": sys_prompt},
        {"role": "user", "content": user_prompt},
    ], add_generation_prompt=True, return_dict=True, return_tensors='pt')


    inp = {f"embedder_{k}": v.to(DEVICE) for k, v in inp.items()}
    output = MODEL.call_embedding_model(**inp)
    inp['frozen_embeddings'] = output["embeddings"]
    with torch.inference_mode():
        out = MODEL.generate(inp, {"max_length": 64})
    inverted = MODEL.tokenizer.decode(out[0], skip_special_tokens=True)
    generated = MODEL.embedder_tokenizer.decode(output["chosen_tokens"][0].squeeze(), skip_special_tokens=True)
    return generated, inverted


demo = gr.Interface(
        fn=invert,
        inputs=gr.Textbox(label="Secret prompt"),
        outputs=(gr.Textbox(label="LLM output"), gr.Textbox(label="Inverter guess"))
        )
demo.launch(share=True)

Citation

@misc{nazir2025betterlanguagemodelinversion,
      title={Better Language Model Inversion by Compactly Representing Next-Token Distributions}, 
      author={Murtaza Nazir and Matthew Finlayson and John X. Morris and Xiang Ren and Swabha Swayamdipta},
      year={2025},
      eprint={2506.17090},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.17090}, 
}

Repository for the paper Better Language Model Inversion by Compactly Representing Next-Token Distributions.

Code: https://github.com/dill-lab/PILS

Downloads last month: -

Safetensors

Model size

7B params

Tensor type

BF16

F32

Paper for dill-lab/pils-32-llama2-chat-7b

Better Language Model Inversion by Compactly Representing Next-Token Distributions

Paper • 2506.17090 • Published Jun 20, 2025 • 1