You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a modified copy of krea/Krea-2-Raw, distributed under the Krea 2 Community License Agreement (LICENSE.pdf in this repository). By agreeing you accept to be bound by its terms and by Krea's Acceptable Use Policy (https://www.krea.ai/krea-2-use-policy).
Log in or Sign Up to review the conditions and access this model content.
Krea 2 Raw IQ8R — radiance container
krea/Krea-2-Raw as a single .rad container for the
radiance inference engine (AMD RDNA4, ROCm). It replaces
StillDeadcode/krea-2-raw-radiance-fp8, with the
same eight bits a weight.
This is not an official Krea product and is not endorsed by Krea.
| File | krea-2-raw-iq8r.rad — 16.37 GiB |
| Weights | the text encoder's (Qwen3-VL-4B), the fusion blocks' and the diffusion transformer's linears int8 (codes −127..127) with a u8 scale per row and 256 of K under a bf16 scale a row, each 128 of K rotated by the unnormalised Hadamard, round to nearest (the fusion's down projection with a bf16 scale per row and 128 instead); the q/k norms and the text norm f32 (as released); the small linears, the remaining norms and tables and the VAE bf16 |
| Activations | int8 under the same rotation (W8A8) |
| Steps | 28 by default, guidance 4.5 against an empty negative prompt (guidance_scale or true_cfg_scale, and negative_prompt); num_inference_steps sets the count |
| Needs | radiance built from main after 1.17.0: IQ8R at one scale a 256 of K is not in 1.17.0 and comes with the next release |
Serve
# radiance.yaml
server:
host: 0.0.0.0
port: 8000
models:
krea-2:
model: krea-2-raw-iq8r.rad
tp: 2
radiance --config radiance.yaml
tp: 1 serves it on one card. The server speaks OpenAI's Images API: POST /v1/images/generations
(Krea 2 reads only text, so /v1/images/edits answers 501). Pictures come back as b64_json (png, jpeg
or webp). Sizes up to 2048x2048, sides multiples of 16, 1024x1024 by default; seed, n, streaming with
partial_images. RGB only (background: transparent is refused).
51 seconds a 1024x1024 picture at 28 guided steps on one Radeon AI PRO R9700 (gfx1201).
Not measured against bf16 at this checkpoint; on Krea 2 Turbo the same recipe is 29.3 dB from the bf16 container's pictures on one card where FP8 is 22.8.
How this file was made
rad-convert krea/Krea-2-Raw --recipe krea-2-iq8r.recipe -o krea-2-raw-iq8r.rad
krea/Krea-2-Raw is the diffusers download of that repository. The recipe (krea-2-iq8r.recipe in this
repository) names what is quantised and cast; everything it does not name is the checkpoint's own dtype
(the release stores some tensors f32, which the casts at the end narrow):
te.blk.*.attn_qkv.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.attn_output.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.ffn_gate_up.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.ffn_down.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
fuse.*.attn_qkv.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
fuse.*.attn_gate.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
fuse.*.attn_output.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
fuse.*.ffn_gate_up.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
fuse.*.ffn_down.weight rtn codes=i8 group=128 scale=bf16 transform=fwht128 clamp=sym rule=search
dit.blk.*.attn_qkv.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
dit.blk.*.attn_gate.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
dit.blk.*.attn_output.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
dit.blk.*.ffn_gate_up.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
dit.blk.*.ffn_down.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
fuse.*.norm1.weight cast dtype=f32
fuse.*.norm2.weight cast dtype=f32
fuse.*.q_norm.weight cast dtype=f32
fuse.*.k_norm.weight cast dtype=f32
dit.blk.*.q_norm.weight cast dtype=f32
dit.blk.*.k_norm.weight cast dtype=f32
txt.norm.weight cast dtype=f32
fuse.* cast dtype=bf16
txt.* cast dtype=bf16
dit.* cast dtype=bf16
vae.* cast dtype=bf16
License
Krea 2 is licensed under the Krea 2 Community License Agreement (LICENSE.pdf); NOTICE carries its
attribution notice. Use must also follow Krea's Acceptable Use Policy,
and the license requires deployers to run content filtering. This container is a modified copy: its weights
are re-encoded as described above. The text encoder inside it is Qwen3-VL-4B, under the Apache License 2.0
(LICENSE-Qwen3-VL).
Model tree for StillDeadcode/krea-2-raw-radiance-iq8r
Base model
krea/Krea-2-Raw