redline-tiny NPU programs
Compiled AMD XDNA2 NPU programs for redline-tiny, which runs Qwen-Image-2.1-Turbo photo editing and text-to-image on the Ryzen AI NPU of a Windows laptop. The installer downloads this repo and checks every file against manifest.json, so there's no need to download it by hand.
| App, installer, kernel sources | GitHub: Cukrus-directive/redline-tiny |
| Gallery, benchmarks, how it works | Space: cukrus/redline-tiny |
| Model weights | Not in this repo. The installer gets the official Qwen/Qwen-Image-2.1-Turbo files and packs them for the NPU on your PC |
What's in it
| Part of the model | Files | Count |
|---|---|---|
| Text encoder (Qwen3-VL-8B: vision tower and language model) | vlm-*/ |
38 programs |
| DiT, step 1 of an edit | merged/dit_step1_v7.elf |
1 merged ELF |
| DiT, steps 2–8 | merged/dit_cached_v7.elf |
1 merged ELF |
| DiT, step 1 of text-to-image | atb3*/, attn32*/, glue-*/, q-atb3k-t2imgin/ |
24 programs, attention once per prompt-length bucket |
| VAE decoder | vae2-*/ |
64 programs |
| Checksums | manifest.json |
size and SHA-256 of every file |
Each program folder holds:
| File | |
|---|---|
final.xclbin |
the program the AIE cores run |
insts.bin |
its instruction stream (data movement, run parameters) |
meta.json |
the design and arguments it was built from, so it can be rebuilt exactly |
final.prj/cdo_main/*.bin |
VAE programs only: the static array setup, used to check that every tensor shape of one kernel shares one program |
A merged ELF holds a whole DiT stage in one hardware context, so the kernels of a step run without context switches. Attention reads its valid-key count at runtime, so one ELF serves every prompt length.
Build
| Toolchain | mlir-aie v1.4.4 (IRON), Peano / llvm-aie 22.0.0.2026092101 |
| Target | npu2: XDNA2 (Strix), 32 AIE cores |
| Tested on | Ryzen AI 9 HX 375 and Ryzen AI 7 350, NPU driver 32.0.20102 |
| Rebuilding | Sources in kernels/, steps in docs/BUILDING.md |
Licence
Apache-2.0. Parts of the GEMM kernels are derived from AMD's mlir-aie examples (Apache-2.0 WITH LLVM-exception). The model these programs run, Qwen-Image-2.1-Turbo, is under the Qwen Research License (non-commercial use only) and is not part of this repo.