FLUX.2 IREE Inference Engine

This repository is a focused implementation path for running black-forest-labs/FLUX.2-klein-4B on Vulkan-first hardware through IREE, with later dynamic LoRA support.

What Is Done So Far

• Built a small, flat export pipeline under model-export-pipeline/.
• Implemented phase 1 quantization: weight-only signed int4 (group size 128) for diffusion and text-encoder weights, with Hugging Face-style output layout.
• Added a quantized model loader that dequantizes at the load boundary, keeping core PyTorch math quantization-agnostic.
• Implemented full phase 2 pure-PyTorch inference (not a smoke probe) and generated validated 1024x1024 images from the quantized model.
• Integrated official FLUX.2 code through the repository submodule at submodules/flux2.
• Cleaned project structure and docs to keep machine-local helpers out of the pushed repo.

Current Direction

The next major step is phase 3: export the working quantized inference path as MLIR/IREE artifacts with fixed initial shapes, then run on the BC-250 Vulkan target.

Near-term priorities:

1. Keep 02_pure_pytorch_inference.py as the behavior reference.
2. Start 03_export_mlir.py from the same quantized model and model code path.
3. Validate first end-to-end IREE execution on Vulkan, then iterate on performance and memory.
4. Add dynamic LoRA loading after the base IREE path is stable.