README.md

---
license: other
license_name: qwen-research
license_link: LICENSE
base_model: abenzerps/Qwen-Image-2.1-Uncensored-GGUF
base_model_relation: quantized
pipeline_tag: text-to-image
tags:
  - mxfp4
  - amd
  - rocm
  - image-editing
  - rgba
---

# Qwen_Image-2.1-Uncensored-MXFP4-Paiton

**Built with Qwen.** This is EliovpAI's balanced MXFP4 conversion of the BF16
checkpoint published by [abenzerps](https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-GGUF),
derived from [Qwen-Image-2.1](https://huggingface.co/Qwen/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](https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-GGUF/discussions/20).
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.

| Runtime | Behavior |
| --- | --- |
| This HF package, compatible `gfx1201` GPU | Automatically uses the included, SHA-256-verified native HIP weight decoder |
| This HF package, other supported ROCm GPU | Uses framework weight reconstruction; tested on MI355X |
| Paiton plugin/container | Uses the existing optimized RDNA4 pipeline, with checkpoint selection at launch |

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:

```bash
./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:

```bash
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.

```bash
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](evaluation/quality-pairs-1.png),
  [remaining pairs](evaluation/quality-pairs-2.png) and
  [complete measurements](evaluation/mi355-correctness.json).
- 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](https://huggingface.co/EliovpAI/Qwen_Image-2.1-MXFP4-Paiton-RDNA4)
and [portable package](https://huggingface.co/EliovpAI/Qwen_Image-2.1-MXFP4) 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](LICENSE), which allows
non-commercial research and evaluation; commercial use requires a separate
upstream license. See [NOTICE](NOTICE) and `artifacts/NOTICE` for attribution and
native-runtime terms.