Skip to content

AGENTS.md

Latest commit

 

History

History
41 lines (30 loc) · 1.76 KB

AGENTS.md

File metadata and controls

41 lines (30 loc) · 1.76 KB
Raw

AGENTS.md

Guidance for AI coding agents working in this repository.

Project overview

thenoise is a focused diffusion inference engine for ROCm (Strix Halo / gfx1151, RDNA 3.5, native BF16/FP16, 128GB unified RAM). It loads one model at a time and exposes a small, explicit surface.

  • Layout:
    • thenoise/ — server package: __main__.py + cli.py (CLI entrypoints), api.py (FastAPI /text2image), runtime.py (single-model runtime), models/ (adapters + catalog + detect), dit/ (per-model compute), vae/ (shared Qwen-Image VAE), utils/ (safetensors, lora, attention, device).
    • scripts/ — model download helpers.
    • tests/ — CLI + runtime + detection tests (no torch needed).
  • Invocation: python -m thenoise serve ... and python -m thenoise generate ....

Critical constraints

  1. Never run uv sync. It would replace/break the ROCm torch build that the maintainer installs directly into the venv. Use uv pip install instead (e.g. uv pip install -e .). torch is intentionally not listed in pyproject.toml.

  2. Never run the program yourself. You are in a containerized environment that cannot run the project with real models — there is no GPU and not enough RAM. Do not attempt to start the server, run generate, or load a model.

  3. When unsure on the correct way to proceed, ALWAYS ask the user rather than guessing or overthinking.

  4. The user is happy to test. Do not be too afraid of breaking stuff, just inform the user of potentially risky changes and ask them to test them out.

Workflow

  • Verify changes with the test suite: .venv/bin/python -m pytest tests/ -q (tests are designed to run without torch or real weights).
  • Do not run heavy/compute-heavy commands.