Skip to content
Other

Drag-and-Drop LLMs

Generate LoRA weights in seconds without training

2 min readupdated

Drag-and-Drop LLMs (DnD) is a method from the authors of a NeurIPS 2025 paper that adapts large language models to new tasks without traditional gradient-based training. Instead of running fine-tuning, the model takes a small batch of unlabeled prompts and maps them directly into a ready-to-use set of LoRA matrices (lightweight trainable adapters layered onto model weights). Under the hood, a lightweight text encoder compresses the prompt batch into condition embeddings, while a cascaded hyper-convolutional decoder synthesizes adapter parameters for a target LLM (large language model) or VLM (vision-language model). It is built for developers and researchers who need to quickly specialize models like Qwen2.5 for reasoning, math, coding, or multimodal tasks.

Key features

  • Training-free adaptation: generates working LoRA weights for a new task in seconds instead of running a full optimization pass.
  • Up to 12,000x lower compute cost: according to the authors, computational overhead is drastically lower than full fine-tuning.
  • Up to 30% quality gains: outperforms top classically trained LoRAs on average across unseen benchmarks in logic, math, coding, and multimodality.
  • Cross-domain generalization: once trained on prompt-checkpoint pairs, the system synthesizes adapters without seeing labeled answers or the target dataset.

What you can do

  • Spin up a task-specific LoRA fast: generate an adapter for Qwen2.5-0.5B/1.5B/7B-Instruct or Qwen2.5-VL-3B-Instruct from just a handful of sample prompts.
  • Reproduce ARC-c experiments: run a ready-made pipeline for common sense reasoning tasks using the authors' released checkpoints.
  • Train your own weight generator: collect custom LoRA checkpoints via LLaMA-Factory, compute layer importance weights, and train DnD for your own domain.

Example

git clone [https://github.com/jerryliang24/Drag-and-Drop-LLMs.git](https://github.com/jerryliang24/Drag-and-Drop-LLMs.git)
cd Drag-and-Drop-LLMs
bash install.sh
hf download Qwen/Qwen2.5-0.5B-Instruct --local-dir models/Qwen/Qwen2.5-0.5B-Instruct
bash scripts/common_sense_reasoning/ARC-c/training_and_generation.sh
Read moreDetails
Paper accepted at NeurIPS 2025 (arXiv: 2506.16406); released weights and logs are available in the Jerrylz/DnD-checkpoints-and-logs dataset.
Supported base models and encoders listed in the docs: Qwen/Qwen2.5-0.5/1.5/7B-Instruct, Qwen/Qwen2.5-VL-3B-Instruct, sentence-transformers/all-MiniLM-L12-v2, and google-t5/t5-base.
Training your own generator requires prepared dataset-to-checkpoint pairs and importance weight calculation via weighted MSE (the example config specifies up to 8 GPUs).
Video example
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Nitin2004/drag-and-drop-llms-nitin-workspace
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Nitin2004/drag-and-drop-llms-nitin-workspace --local-dir ~/models/drag-and-drop-llms-nitin-workspace
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of drag-and-drop-llms-nitin-workspace (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

Nitin2004/drag-and-drop-llms-nitin-workspace
Author
Nitin2004
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments