duo-laguna-adapter-smoke

DuoAttention Laguna Adapter

This repository contains learned DuoAttention attention-head weights and custom loading code for poolside/Laguna-XS.2. It intentionally does not include the full Laguna base-model weights or tokenizer files.

Install optional tokenizer dependencies if needed:

pip install sentencepiece tiktoken

Load the tokenizer from the base Laguna model and the patched model from this adapter repository:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

adapter_repo = "dogeplusplus/duo-laguna-adapter-smoke"
base_model = "poolside/Laguna-XS.2"

tokenizer = AutoTokenizer.from_pretrained(
    base_model,
    trust_remote_code=True,
    token=True,
)

model = AutoModelForCausalLM.from_pretrained(
    adapter_repo,
    trust_remote_code=True,
    token=True,
    torch_dtype="auto",
    device_map="auto",
)

prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    generated = model.generate(**inputs, max_new_tokens=32)
print(tokenizer.decode(generated[0], skip_special_tokens=True))

Use token=True after running hf auth login, or pass a token string directly when loading private or gated repositories.