aitrepreneur-docker

KO | EN

gitlite — search

#python #java #python3 #arduino #golang #machine-learning #rust #html #flask #javascript #seismology #nodejs

aitrepreneur-docker

★ 17 Open GitHub ↗

Personal repo from aitrepreneur's recipes in docker automated format

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Aitrepreneur RunPod stacks — local Docker (NVIDIA GPU)

Runs the tools from [Aitrepreneur](https://www.youtube.com/@Aitrepreneur)'s
RunPod one-click scripts locally, as reproducible Docker images: pinned
versions, persistent data on the host/NAS, one command to run or upgrade each
app.

All credit for the stack selection, model curation, version pinning research,
and workflows goes to Aitrepreneur — this repo only translates his work into
Docker form. His scripts, workflows, and templates are **paid Patreon content**
and are **not** included here: if you find this useful, please support him at
**<https://www.patreon.com/c/aitrepreneur/home>** — that's also where you get
the original files. See [`docs/reference/README.md`](docs/reference/README.md)
for where to place them if you have access (only the Ideogram workflow JSON is
actually used by the build; the rest is optional reference).

| Service | What it is | Start | UI |
|---|---|---|---|
| `ai-toolkit` | [ostris/ai-toolkit](https://github.com/ostris/ai-toolkit) LoRA training UI | `make up` | <http://localhost:8675> |
| `krea2` | ComfyUI + Krea 2 image generation | `make krea2` | <http://localhost:8188> |
| `ideogram` | ComfyUI + Ideogram 4 (typography/design) | `make ideogram` | <http://localhost:8189> |
| `ltx` | ComfyUI + LTX-2.3 video generation | `make ltx` | <http://localhost:8190> |

Ports are the defaults; override them in `.env` (`AI_TOOLKIT_PORT`,
`KREA2_PORT`, `IDEOGRAM_PORT`, `LTX_PORT`).

Each app gets its own container because their Python stacks conflict
(different torch/transformers pins). They share the NAS models tree, so a
model downloaded once is available to all.

## Prerequisites

- NVIDIA driver + nvidia-container-toolkit (verify with `make gpu-check`).
  Compose uses the legacy `nvidia` runtime rather than `gpus: all` because the
  latter reads the CDI spec (`/etc/cdi/nvidia.yaml`), which breaks after
  driver updates until regenerated
  (`sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml`).
- Docker with Compose v2; ~20 GB image disk per app.
- The NAS mounted at `/mnt/gigachad` (models land there — LTX alone is ~40 GB).
- One RTX 5090 = run one heavy app at a time (`make stop-krea2` before
  `make ltx`, etc.). The containers themselves can coexist.

## Quick start

```bash
git clone git@github.com:akitaonrails/aitrepreneur-docker.git
cd aitrepreneur-docker
make setup          # creates .env — edit it to add your HF_TOKEN
make build          # builds all four images (long first time)
make up             # ai-toolkit at http://localhost:8675
make krea2          # ComfyUI+Krea2 at http://localhost:8188
```

Model files are **not** in the images: each ComfyUI app checks
`comfyui/apps/<app>/models.txt` at container start and downloads only what is
missing from the NAS models tree. First start of an app therefore takes a
while — watch with `make logs-krea2`.

Day to day: `make up` / `make krea2` / `make ideogram` / `make ltx`,
`make stop-<app>`, `make logs-<app>`, `make shell-<app>`, `make down`.

## Data layout

Everything stateful lives outside the containers. Rebuilding or upgrading
never touches data, and nothing big can land in a container's writable layer
(all caches are env-redirected onto volumes).

```
./data/<app>/                            # small, local (gitignored)
    output/, input/, user/    ComfyUI apps: results, uploads, UI state
    datasets/, outputs/, db/  ai-toolkit: training data, LoRAs, job db

/mnt/gigachad/comfyui/models/            # big, on the NAS
    diffusion_models/, text_encoders/, vae/, loras/, unet/, ...
                 # shared ComfyUI-layout tree: your existing models plus
                 # whatever the apps download (skip-if-exists)
    hf-cache/    # ai-toolkit's Hugging Face cache
    aitk-cache/  # torch.hub + misc library caches, shared by all apps

/mnt/gigachad/comfyui/ideogram-templates/  # extracted reference templates
```

The `ideogram` container also seeds the bundled Ultra workflow into its UI
(`comfyui/apps/ideogram/workflows/`) and mounts the template images read-only
at `input/templates`. ai-toolkit mounts the NAS models tree read-only at
`/comfyui-models` so training configs can reference existing checkpoints.

## How an app is defined

Each ComfyUI app is a directory under `comfyui/apps/<name>/` — that's the
whole upgrade surface:

| File | Purpose |
|---|---|
| `nodes.txt` | custom nodes baked into the image (`folder url [git-ref]`) |
| `constraints.txt` | pip pins protected during every install |
| `extras.txt` | extra pip packages the workflows need |
| `models.txt` | model files fetched at container start (skip-if-exists) |
| `check.sh` | optional build-time sanity check |
| `workflows/` | workflow JSONs seeded into the UI |

Add a node → append to `nodes.txt`, `make upgrade-<app>`. New model file →
append to `models.txt`, restart the app. ComfyUI-Manager works in the UI for
experiments, but anything it installs lives in the container layer and
disappears on rebuild — promote keepers into `nodes.txt`. A pip blacklist is
seeded into each app's user dir so the Manager can't replace the locked
torch/transformers stack (same protection as the reference scripts).

## Upgrading

```bash
make upgrade-krea2   # rebuild one app from fresh upstream clones
make upgrade         # rebuild everything
make version         # upstream commits baked into each image
```

See [`docs/UPGRADING.md`](docs/UPGRADING.md) for version-pin locations and
the torch/CUDA compatibility notes.

## Deviations from the reference scripts

- **torch cu128 everywhere.** The ideogram/ltx scripts pin torch 2.4.0+cu121,
  which has no RTX 50xx (Blackwell/sm_120) kernels and cannot run on this
  machine's GPU. All ComfyUI apps here use torch 2.8.0+cu128 (the krea2
  script's own pin); xformers is dropped for ltx (cu121-only build) in favor
  of ComfyUI's native PyTorch attention.
- **Install at build, not at boot.** The scripts install apt/pip/nodes on
  every fresh pod; here that's baked into images, and pods'/containers' state
  can't drift — rebuilding is the only way anything changes.
- **Models on the NAS, atomically.** Downloads go to `.part` files and are
  renamed only when complete, so an interrupted download is never mistaken
  for a finished model (the scripts' skip-if-exists check has that flaw).
- **krea2 node fix.** The reference script clones `ComfyUI-Krea2T-Enhancer`
  but forgot to list it in `REQUIRED_NODES`, so its Python requirements were
  never installed. Here every cloned node gets its requirements.

aitrepreneur-docker

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README