EngineAI Studio Pro

Kimodo demo on Linux + NVIDIA GPU: text-to-motion generation, NF4 text encoder (matbee/kimodo-llm2vec-nf4), optional retargeting to the EngineAI T800 humanoid.

Model weights are not in the repository — download them with the scripts below after cloning.

Quick start

git clone https://github.com/robofighter/engineAI-studio-pro.git
cd engineAI-studio-pro

sudo ./install_system_deps.sh    # once per machine
./install.sh
source ./activate_cuda.sh
./scripts/verify_gpu_setup.sh

./download_nf4.sh
hf auth login                     # only for nvidia/Kimodo-*
./scripts/download_kimodo_models.sh

./run_demo.sh                     # http://127.0.0.1:7860

Full step-by-step setup, WSL2, RTX 50xx cards, and common errors — docs/INSTALL.md.

Hardware

Tested on: NVIDIA RTX 50xx (Blackwell), 12 GB VRAM, Linux / WSL2.

VRAM in practice: single-process run_demo.sh (NF4 encoder + diffusion) uses about 7–7.5 GB on GPU. It may work on 8 GB cards in theory, but that has not been fully validated — leave headroom or use the two-process split below.

• 12 GB — comfortable for ./run_demo.sh (one process)
• ~8 GB — try ./run_demo.sh first; if OOM, use ./run_textencoder.sh + ./run_demo_api.sh (see docs/INSTALL.md)
• 16 GB+ — same as 12 GB, more margin for multi-sample / T800
• Ubuntu 22.04/24.04 or Debian 12, x86_64 (WSL2 works)
• ~40–50 GB disk (conda, PyTorch, models)
• 16 GB system RAM

In git vs downloaded after clone

More docs

• docs/INSTALL.md — main installation guide
• docs/GPU.md — GPU vs CPU, environment variables
• docs/MAINTAINER.md — publishing to GitHub
• .env.example — optional environment overrides

Publish to GitHub

cp .env.github.example .env.github   # set GITHUB_USER and GITHUB_TOKEN
./scripts/publish_to_github.sh

Do not commit .env.github (it is gitignored).

Licenses

Kimodo code — Apache 2.0 (see kimodo/LICENSE and LICENSE). NVIDIA / Meta / matbee models — Hugging Face terms. SMPL-X requires separate registration on the SMPL-X website.