Llama-3-Kinisis-LLM2Vec-NF4

Llama-3-Kinisis-LLM2Vec-NF4

Built with Meta Llama 3.

A 4-bit (NF4) text encoder for the NVIDIA Kimodo text-to-motion model, sized for consumer hardware (~5.4 GB on disk; runs on a 10 GB+ GPU, or CPU-offloaded next to a smaller GPU). It is the text-conditioning encoder used by Kinisis Motion, a prompt-driven animation plugin for Unreal Engine.

What it is

• Base: meta-llama/Meta-Llama-3-8B-Instruct
• Merged in (baked, no runtime PEFT): the LLM2Vec bidirectional-encoder adapters McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp and McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised
• Quantization: bitsandbytes NF4, double quantization, fp16 compute (quantization_config is serialized in config.json — from_pretrained restores it automatically; requires bitsandbytes)
• Architecture: LlamaBiModel (bidirectional attention, LLM2Vec), mean pooling, max length 512 (llm2vec_config.json)
• Output: 4096-dim text embeddings

<!-- Fill in at publish time (BAKE_RUNBOOK.md): -->

• Baked on: <date>, base revision <meta-llama revision>, adapter revisions <mntp revision> / <supervised revision>, tooling: Aero-Ex/kimodo fork commit bfc2dc4 (Apache-2.0)

How to use

Point the Kinisis Motion local-server installer at this repo (it does this by default) — or load it directly:

from kimodo.model.llm2vec.llm2vec import LLM2Vec  # nv-tlabs/kimodo
model = LLM2Vec.from_pretrained(
    base_model_name_or_path="<this repo or a local download>",
    peft_model_name_or_path=None,   # adapters are already baked in
    torch_dtype=torch.float16,
    device_map={"": "cuda:0"},      # or {"": "cpu"} to hold it in system RAM
)
emb = model.encode(["A person walks forward"], batch_size=1)

License and attribution

This model is a derivative of Meta Llama 3 and is distributed under the Meta Llama 3 Community License (see LICENSE); use is subject to Meta's Acceptable Use Policy. See NOTICE for the required attribution.

• Meta Llama 3: Copyright © Meta Platforms, Inc. All Rights Reserved.
• LLM2Vec (McGill NLP Group): adapters published on Hugging Face; code MIT — https://github.com/McGill-NLP/llm2vec
• Kimodo (NVIDIA): Apache-2.0 code; motion-model weights under the NVIDIA Open Model License (not included here — this repo is only the text encoder)
• NF4 bake tooling: Aero-Ex/kimodo fork, Apache-2.0

This repository is published by Elena Dimopoulou for the Kinisis Motion product and is not affiliated with or endorsed by Meta, NVIDIA, or McGill NLP.