StillDeadcode/radiance-qwen-image-2.1-fp8

Qwen-Image 2.1, FP8 — radiance container

Qwen/Qwen-Image-2.1 — text to image and image editing — as a single .rad container for the radiance inference engine (AMD RDNA4, ROCm): the Qwen3-VL-8B conditioning network with its vision tower, the 7.1B diffusion transformer and the RGBA VAE.

File	qwen-image-2.1-fp8.rad — 16.08 GiB
Weights	the conditioning network's (Qwen3-VL-8B) and the diffusion transformer's linears FP8 E4M3 per 128×128 block (bf16 scales); the vision tower, the small linears, the norms and the VAE bf16
Steps	40 by default (num_inference_steps, or quality: low / medium / high)
Needs	radiance built from main after 1.5.0 — Qwen-Image support is not in 1.5.0 and comes with the next release

Serve

# one card
radiance --model qwen-image-2.1-fp8.rad --host 0.0.0.0 --port 8000
# two cards; --tp-wire wht6 sends the all-reduces as 6-bit (about 9% faster, within a dB of exact)
radiance --model qwen-image-2.1-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, streaming with partial_images, background: transparent; edits take 1 to 10 reference pictures and a mask. /v1/models lists the model.

Seconds a picture on Radeon AI PRO R9700 (gfx1201); 2048x2048 at the default 40 steps, 1024x1024 at 20:

	2048x2048	1024x1024	1024x1024 edit
one card	104.9	9.5	11.9
two cards	61.1	6.0	7.4
two cards, --tp-wire wht6	55.8	5.2	6.6

Against diffusers' own bf16 pipeline from the same noise (1024x1024, 20 steps), PSNR 28.9 dB on a prompt and 38.5 dB on an edit; radiance's bf16 build of the same model is at 33.9 / 42.2 dB, where diffusers differs from itself by 32.6 to 34.4 dB across its attention kernels.

How this file was made

rad-convert Qwen/Qwen-Image-2.1 --recipe qwen-image-2.1-fp8.recipe -o qwen-image-2.1-fp8.rad

Qwen/Qwen-Image-2.1 is the diffusers download of that repository.

The recipe (qwen-image-2.1-fp8.recipe in this repository) 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 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 linear weights are re-encoded as FP8 (see above).