LoRAForge

Train LoRAs on the GPU you actually own.

(Working title — rename freely.)

A hardware-aware, recipe-driven LoRA training studio for NVIDIA gaming cards. One-click install on Windows and Linux, no WSL required. Inspired by the recipe architecture of NVIDIA NeMo AutoModel, sized for a single consumer GPU.

Why another trainer?

Every existing tool fails users the same three ways: brutal installation, walls of unexplained parameters, and configs silently copied from a 4090 tutorial onto an 8GB laptop card. LoRAForge attacks exactly those three:

1. Diagnose first. On startup we probe your GPU, VRAM, driver, torch/CUDA build, RAM, and disk — then tell you what you can train before you waste a minute. Unavailable options are disabled with a reason, never hidden.
2. Recipes, not knob soup. Every run is a validated YAML recipe resolved against a community-maintained capability matrix (model × VRAM tier × GPU architecture). Known-good presets per card; an advanced pane for tinkerers, pre-filled with safe values.
3. Install that just works. A small desktop shell bootstraps a pinned Python environment with uv — including the correct torch wheel line for your GPU generation (cu126 / cu128+ for Blackwell). First run in about a minute, not an afternoon.

Architecture

┌────────────────────────────────────────────────┐
│  Desktop shell (Tauri) / browser → localhost   │
├────────────────────────────────────────────────┤
│  FastAPI job server  (this package)            │
│    probe.py        hardware diagnostic         │
│    capability/     matrix.yaml + resolver      │
│    recipes/        pydantic recipe schema      │
│    engines/        adapter protocol            │
├────────────────────────────────────────────────┤
│  Engine environments (isolated, uv-managed)    │
│    kohya sd-scripts   ← default (Apache-2.0)   │
│    musubi-tuner       ← video (later)          │
│    SimpleTuner        ← optional ext. service  │
└────────────────────────────────────────────────┘

Principles borrowed from NeMo AutoModel:

• The recipe is the single source of truth. Shareable, reproducible, overridable from the CLI/UI. Imagine 4070-flux-character.yaml shared on Civitai the way presets are shared today.
• Preprocessing is a separate stage. VAE latents and text embeddings are cached to disk before training; the training step loads only the transformer. This is the single biggest VRAM lever available.
• Capability data, not capability code. What fits on which card is a versioned YAML file the community can extend with PRs — not logic buried in UI code.
• Typed configs. Validation errors read like "peft.alpha must be a number", not a stack trace 40 seconds into a run.

Status

Pre-alpha skeleton. The probe, capability resolver, recipe schema, and engine adapter protocol are real and tested; the server, UI, and kohya adapter are next.

Development

uv venv && uv sync --extra dev
uv run pytest
uv run loraforge diagnose        # prints your hardware capability report

Cross-platform rules from the first commit: pathlib everywhere, no Unix-only shell-outs, CI on Linux and Windows, and the happy path never requires flash-attn or Triton (SDPA works everywhere; compile-based speedups are detected extras).

License

Apache-2.0.