recursal/QRWKV6-7B-Instruct

RADLADS: Rapid Attention Distillation to Linear Attention Decoders at Scale

This repository contains the model described in the paper RADLADS: Rapid Attention Distillation to Linear Attention Decoders at Scale.

RADLADS (Rapid Attention Distillation to Linear Attention Decoders at Scale) presents a novel protocol for rapidly converting softmax attention transformers into linear attention decoder models. This innovative process requires only 350-700 million tokens for distillation, which is less than 0.005% of the tokens used to train the original teacher models. Despite this minimal training, the inference quality remains remarkably close to the original transformer.

These models achieve state-of-the-art downstream performance across a set of standard benchmarks for linear attention models of their size, offering significant efficiency benefits with constant-time inference per token. The project also introduces two new RWKV-variant architectures, RAD-RWKV6 and RAD-RWKV7, which serve as efficient destination architectures for transformer conversions.

We release all our models on Hugging Face under the Apache 2.0 license. Please note that our 72B models are also governed by the Qwen License Agreement.

Github repository: https://github.com/recursal/Monet

Quickstart

You can explore the core implementation of RADLADS in the GitHub repository. To use these models with the Hugging Face transformers library, you will need to set trust_remote_code=True when loading them due to custom architecture components.

Text Generation

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Replace with the actual model ID (e.g., recursal/radrwkv7qwen2-7b-instruct)
model_id = "your-model-id-here" 

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
).eval()

text = "The quick brown fox jumps over the lazy"
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=20, do_sample=True, temperature=0.7, top_p=0.8, top_k=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Chat Completion

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Replace with the actual model ID (e.g., recursal/radrwkv7qwen2-7b-instruct)
model_id = "your-model-id-here" 

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
).eval()

messages = [
    {"role": "system", "content": "You are a helpful AI assistant."},
    {"role": "user", "content": "What is the capital of France?"}
]

# Apply chat template and generate text
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50, do_sample=True, temperature=0.7, top_p=0.8, top_k=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

If you use this code or find our work valuable, please consider citing RADLADS:

@misc{goldstein2025radladsrapidattentiondistillation,
      title={RADLADS: Rapid Attention Distillation to Linear Attention Decoders at Scale}, 
      author={Daniel Goldstein and Eric Alcaide and Janna Lu and Eugene Cheah},
      year={2025},
      eprint={2505.03005},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.03005}, 
}