Krea-2-Turbo-FP8-SGLang

Krea-2-Turbo-FP8-SGLang

SGLang-ready FP8 transformer weights for krea/Krea-2-Turbo.

This repo contains only the converted Krea2 transformer. It does not include the Krea2 text encoder, VAE, tokenizer, scheduler, or LoRA weights. Use those components from krea/Krea-2-Turbo, and use this repo as the --transformer-weights-path / TRANSFORMER_WEIGHTS_PATH override in SGLang Diffusion.

Provenance

• Base model: krea/Krea-2-Turbo
• FP8 source: sakamakismile/Krea-2-Turbo-FP8
• Conversion target: SGLang Diffusion Krea2 transformer format
• Quantization metadata: ModelOpt FP8 with weight_scale and input_scale tensors
• License: inherits the krea-2-community-license from Krea's base model

This is a community conversion. It is not affiliated with Krea or SGLang.

Files

LICENSE.pdf
config.json
diffusion_pytorch_model-00001-of-00002.safetensors
diffusion_pytorch_model-00002-of-00002.safetensors
diffusion_pytorch_model.safetensors.index.json
UPLOAD_MANIFEST.tsv

The uploaded checkpoint intentionally excludes modelopt_state.pth. The goal of this repository is the SGLang-loadable transformer form, not the original ModelOpt restore form.

Usage

Example server launch:

export MODEL_PATH=$HOME/models/krea/Krea-2-Turbo
export TRANSFORMER_WEIGHTS_PATH=$HOME/models/krea/Krea-2-Turbo-FP8-SGLang
export ENABLE_CACHE_DIT=1
export BACKEND=sglang

CUDA_VISIBLE_DEVICES=0 \
SGLANG_CACHE_DIT_ENABLED=true \
sglang serve \
    --model-path "$MODEL_PATH" \
    --transformer-weights-path "$TRANSFORMER_WEIGHTS_PATH" \
    --num-gpus 1 \
    --host 127.0.0.1 \
    --port 30100 \
    --backend sglang

The runtime must support Krea2 ModelOpt FP8 loading, including quant_config, weight_scale, and input_scale handling in the Krea2 transformer. If the server reports a roughly BF16-size resident transformer or logs skipped scale keys, the FP8 path is not active.

LoRA

Runtime LoRA was tested with krea/Krea-2-LoRA-retroanime.

For this FP8 transformer, use dynamic LoRA rather than merging LoRA weights into the resident FP8 weights:

{
  "lora_nickname": "test_lora",
  "lora_path": "/path/to/Krea-2-LoRA-retroanime",
  "strength": 0.8,
  "target": "transformer",
  "merge_mode": "dynamic"
}

Direct merge is not compatible with the transposed FP8 resident layout used in the tested SGLang path.

Local Performance

Environment:

• Ubuntu 24.04.4
• NVIDIA GeForce RTX 4090
• SGLang source build with native Krea2 backend
• Krea2 image size: 1024x1024
• Inference steps: 8
• Concurrency: 1

The generic Diffusers fallback path on the same host was much slower, around 58-60 seconds per 1024px image. If you see that behavior, you are probably not running the native SGLang Krea2 backend.

Quality Check

A local BF16-vs-FP8 comparison used the same prompts, 1024x1024 size, 8 steps, and seeds [12345, 12346, 12347] by script. The earlier BF16 metadata was inferred from output timestamps and script order, so this is a practical reproducibility check rather than a perfect lab-grade A/B record.

Across 12 paired images:

• Average SSIM: 0.9246
• Average PSNR: 21.92 dB
• Average MAE: 11.16
• Average histogram cosine: 0.9853
• Average dHash similarity: 0.9349

Visual inspection found no blank frames, severe color shift, stripe artifacts, tiling, or FP8-specific collapse. Differences were concentrated in high-frequency details such as hair edges, background texture, propellers, and LoRA-stylized anime details.

Caveats

• This is a transformer-only repository.
• SGLang must support Krea2 ModelOpt FP8 scale tensors for this to load as FP8.
• Dynamic LoRA is recommended; direct LoRA merge is not supported in the tested FP8 path.
• Performance numbers are single-host local measurements, not a universal benchmark.

License

This repository inherits the krea-2-community-license from krea/Krea-2-Turbo. The original Krea license PDF is included as LICENSE.pdf. Read and follow it before using or redistributing these weights.