Skip to content
Other

UniGFM-RAG

end-to-end graph retrieval and reasoning

2 min readupdated

A model checkpoint from the IEEE TKDE 2026 paper of the same name that unifies Graph Foundation Models (GFMs) and large language models (LLMs) into a single GraphRAG framework. It handles complex multi-hop retrieval and reasoning over knowledge graphs, generating answers directly from discovered relationships. Under the hood, a GFM-based retriever connects to an LLM-based generator via a differentiable graph bridge, enabling joint training of both retrieval and generation. Useful for researchers and engineers working with graph knowledge bases and complex question-answering systems.

Key features

  • Unifies Graph Foundation Models (GFMs) and LLMs into a single end-to-end GraphRAG pipeline.
  • Supports complex multi-hop retrieval and reasoning.
  • Connects the retriever and generator through a differentiable graph bridge.
  • Jointly optimizes data retrieval and answer generation end-to-end.

What you can do

  • Reproduce the GraphRAG experiments from the IEEE TKDE 2026 paper.
  • Use the checkpoint as a baseline for research on multi-hop question-answering (QA) systems.
  • Integrate the architecture into research prototypes for retrieval and reasoning over structured graphs.

Example

The model card does not include ready-to-run inference code, but outlines the core research workflow: feeding a complex multi-hop query against a knowledge graph, where the GFM retriever passes relevant context through the graph bridge to the LLM to generate the final answer. The authors refer readers to the arXiv paper for implementation details.

Read moreDetails
Publication: IEEE Transactions on Knowledge and Data Engineering (TKDE), 2026
Task: graph-ml, end-to-end GraphRAG, multi-hop QA
Limitations: the README omits details on the base LLM, parameter count, VRAM requirements, and code examples (these are covered in the paper)
Discuss Telegram 1

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 46 GB · ready files51 GB
  • FP8 / INT8recommendedweights 23 GB26 GB
  • 4-bitweights 13 GB15 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

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.

    # cheng11/UniGFM-RAG
    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 cheng11/UniGFM-RAG --local-dir ~/models/UniGFM-RAG
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to UniGFM-RAG
  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 UniGFM-RAG (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

cheng11/UniGFM-RAG
Author
cheng11
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments