license: other
license_name: qwen-research
license_link: LICENSE
pipeline_tag: text-to-image
library_name: radiance
tags:

• radiance
• rocm
• amd
• rdna4
• fp8
• text-to-image
• image-editing
• uncensored
• not-for-all-audiences
  base_model:
• Qwen/Qwen-Image-2.1-Turbo
• AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF

Qwen-Image 2.1 Turbo with AtomicChat's abliterated text encoder, FP8 — radiance container

Qwen-Image 2.1 Turbo with the abliterated Qwen3-VL-8B text encoder
of AtomicChat's Qwen-Image-2.1-Turbo-Uncensored,
as a single .rad container for the radiance inference engine
(AMD RDNA4, ROCm).

What differs from Qwen-Image 2.1 Turbo

Only the conditioning network's language model: the o_proj and down_proj weights of all 36 layers and the token
embedding, which AtomicChat's abliteration rewrote (scale 0.75, per their GGUF's metadata). The diffusion transformer,
the vision tower and the VAE are the stock Turbo model's.

What that does to pictures is small. The stock pipeline has no content filter, and AtomicChat's own test found no
image outcome changed over 45 prompts; here a picture moves by about as much as the FP8 conversion moves it (39.7 dB
PSNR from the stock container at 1024x1024). The Turbo diffusion transformer itself has no uncensored fine-tune on
Hugging Face.

Serve

radiance --model qwen-image-2.1-turbo-uncensored-fp8.rad --tp 2 --tp-wire wht6 --host 0.0.0.0 --port 8000

The server speaks OpenAI's Images API: POST /v1/images/generations and POST /v1/images/edits
(JSON, or multipart with image / image[] and mask, as OpenAI's SDKs send them). Pictures come back as
b64_json (png, jpeg or webp; response_format: url is refused). Also: n, size (WxH or auto =
2048x2048), seed, negative_prompt with true_cfg_scale, streaming with partial_images,
background: transparent; edits take 1 to 10 reference pictures and a mask.

On two Radeon AI PRO R9700 with --tp-wire wht6: 11.6 s at 2048x2048, 2.2 s at 1024x1024, 3.1 s for a 1024x1024
edit -- the stock Turbo model's speed.

How this file was made

python extract_atomic.py Qwen/Qwen-Image-2.1/text_encoder AtomicChat/Qwen-Image-2.1-Turbo-Uncensored/text_encoder
rad-convert AtomicChat/Qwen-Image-2.1-Turbo-Uncensored --recipe qwen-image-2.1-fp8.recipe -o qwen-image-2.1-turbo-uncensored-fp8.rad

extract_atomic.py (in this repository) reads AtomicChat's BF16 GGUF by HTTP range and writes the 73 tensors that
differ from the stock text encoder as abliterated.safetensors, with a model.safetensors.index.json that points
those names at it and every other name at the stock shards (linked beside it). AtomicChat/Qwen-Image-2.1-Turbo-Uncensored
is the diffusers download of Qwen/Qwen-Image-2.1-Turbo with that text_encoder/ in place.

The recipe (qwen-image-2.1-fp8.recipe in this repository) is the stock containers' — it names what is
quantised; everything it does not name is the checkpoint's own dtype:

te.blk.*.attn_qkv.weight             rtn  codes=fp8_e4m3 block=128x128 scale=bf16
te.blk.*.attn_output.weight          rtn  codes=fp8_e4m3 block=128x128 scale=bf16
te.blk.*.ffn_gate_up.weight          rtn  codes=fp8_e4m3 block=128x128 scale=bf16
te.blk.*.ffn_down.weight             rtn  codes=fp8_e4m3 block=128x128 scale=bf16
dit.blk.*.attn_qkv.weight            rtn  codes=fp8_e4m3 block=128x128 scale=bf16
dit.blk.*.attn_output.weight         rtn  codes=fp8_e4m3 block=128x128 scale=bf16
dit.blk.*.ffn_gate_up.weight         rtn  codes=fp8_e4m3 block=128x128 scale=bf16
dit.blk.*.ffn_down.weight            rtn  codes=fp8_e4m3 block=128x128 scale=bf16
vae.*                                cast dtype=bf16

License

Built with Qwen. Qwen-Image 2.1 Turbo is licensed under the Qwen RESEARCH LICENSE AGREEMENT (non-commercial;
LICENSE in this repository), and NOTICE carries its attribution. This container is a modified copy: its text
encoder carries AtomicChat's abliteration and its linear weights are re-encoded as FP8 (see above).