EliovpAI/Qwen_Image-2.1-Uncensored-MXFP4-Paiton

Qwen_Image-2.1-Uncensored-MXFP4-Paiton

Built with Qwen. This is EliovpAI's balanced MXFP4 conversion of the BF16 checkpoint published by abenzerps, derived from Qwen-Image-2.1. Fine-tune credit belongs to abenzerps; Paiton supplies the conversion and runtime integration. This was converted from BF16, not from a lower-bit GGUF quant.

The author describes a merged LoRA in the source discussion. The training recipe is not published. “Uncensored” retains the author's model label; our ordinary-image correctness screen does not verify every claimed fine-tuning behavior or guarantee unrestricted outputs.

Weights and runtimes

Tensor payload: 8.674 GiB. Storage uses E2M1 MXFP4 with one E8M0 scale per 32 columns. The balanced recipe packs 224 transformer and 372 text/vision targets. Transformer input/output/conditioning, normalizations, biases and the VAE retain BF16 inference precision. Activations remain BF16 in this package.

The supplied loader is required; an unmodified DiffusionPipeline.from_pretrained cannot load this packed format.

The included decoder is the existing released binary, unchanged. It has no Torch or Triton dependency; the surrounding image pipeline uses the pinned framework. The additional attention, normalization and GEMM optimizations are supplied by the plugin/container. The compiler and generated implementation source are not included. The portable decoder does not provide those serving optimizations.

One container, either model

The plugin's v1.0.3 launcher selects the model before loading it:

./models/Qwen-Image-2.1/serve-docker.sh --model uncensored
./models/Qwen-Image-2.1/serve-docker.sh --model original

Run one worker at a time. original remains the default. Both selections use the same container and native libraries, with separate pinned HF revisions and file hashes. The uncensored selection defaults to the exact BF16 arithmetic profile; the original retains its existing schedule-int8 default. Low-precision profiles remain explicit options and have not been graded on this fine-tune.

Without a checkout:

docker run --rm --device /dev/kfd --device /dev/dri --ipc=host -p 127.0.0.1:8191:8191 -v paiton-qwen-image21-cache:/cache ghcr.io/eliovp/paiton-vllm-plugin:qwen-image21-mxfp4-rdna4-v1.0.3 --model uncensored

The API identifies this checkpoint as paiton-image-2.1-uncensored. A generation request may omit model to use the loaded checkpoint. Switching checkpoints requires restarting the worker; two models are not resident simultaneously.

Direct package use

Download this repository with the Hugging Face CLI. Use Python 3.12 and a ROCm PyTorch stack compatible with the GPU. Install the pinned image dependencies from requirements.txt; requirements-mi355.lock.txt records the tested MI355 environment. Native decoder use requires a runtime able to load the included ROCm 10 artifact; an unavailable native library falls back to framework decoding in auto mode and records the reason.

hf download EliovpAI/Qwen_Image-2.1-Uncensored-MXFP4-Paiton --local-dir model
cd model
python -m pip install -r requirements.txt
./launch.sh --model-dir . generate --prompt 'A green ceramic teapot on a wooden table' --size 2048 --seed 44 --output ../outputs/teapot.png

Use --decoder framework or --decoder native before generate to choose explicitly. --mode rgba requests transparency. Editing uses --mode edit, one --image and --size 1024. Supported generation sizes are 1024 and 2048 square, batch one, 40 steps and guidance 1.0. No CPU model offload or VAE tiling is used. Whole-pipeline graph capture is unsupported.

Correctness and measured limits

Conversion and correctness were tested on MI355X (`gfx950`), Torch 2.10.0+rocm7.1, HIP 7.1.25424, Quark 0.12.post1+rocm71.torch2.10, and Diffusers commit 7263f3317f6b392d62f41e9d75ed9d7e21fc5a5c.

• —All 297 BF16 source tensors were extracted losslessly. All 224 converted transformer layers matched Quark through an independent MXFP4 decoder, and all 73 intentionally BF16 transformer tensors remained exact.
• —Eight paired 2048-square images were compared with the fine-tune's BF16 reference. Mean CLIPScore was 0.9024 → 0.8688; mean LPIPS was 0.1927. Quantization can change composition and text rendering. This is a descriptive screen, not an image-equivalence or human-preference guarantee. See the paired images, remaining pairs and complete measurements.
• —Non-default streams, RGBA with alpha spanning 0–255, editing and A–B–A passed. Repeated prompt embeddings and final denoiser latents were bit-identical. PNG pixels showed the same small VAE variability seen in BF16 controls: repeated-image RGB PSNR was approximately 61.2 dB.
• —Instrumented MI355 MXFP4 generation took 23.60 seconds for the first tested 2048 request and 21.13–21.52 seconds across seven subsequent, different prompts. Sampled whole-device peak was 32.31 GiB for that screen and 34.59 GiB across the mode suite. A 30 GiB framework allocator cap failed during VAE decoding; successful correctness runs used a 64 GiB cap. These are instrumented framework results, not R9700 serving measurements.

This fine-tune has not been run on R9700 in this qualification. Its tensor layout matches the existing optimized pipeline, but the original model's R9700 latency, VRAM and low-precision quality measurements do not qualify this new checkpoint. Its native decoder is reused from the previously qualified release.

Provenance and license

Source revision: 40319fb15542f0ad22921e0124a191a8a935a60a. Source file: qwen-image-2.1-UC-BF16.gguf. SHA-256: f151c683a8aed4b310777017ebbbe3f2180f1180f7867115171adb7d50b0762a. The unchanged original text encoder and VAE come from Qwen revision 790c92633540aa0cb11d9abf19eb46d861714758; the encoder reuses our verified balanced conversion. result.json and checkpoint.lock.json record the complete layout and hashes. Conversion tools are included under tools/ and require the pinned MI355 Quark environment; inference does not require Quark.

The original RDNA4 package and portable package contain the same original-model tensor data, with different loaders and metadata. This repository contains the distinct fine-tune and keeps the Paiton storage format for plugin compatibility.

The weights remain subject to the Qwen Research License, which allows non-commercial research and evaluation; commercial use requires a separate upstream license. See NOTICE and artifacts/NOTICE for attribution and native-runtime terms.