BBuf/flux1-dev-modelopt-nvfp4-sglang-transformer

FLUX.1-dev ModelOpt NVFP4 SGLang Transformer Override

This repository contains a mixed BF16 plus NVFP4 transformer override for black-forest-labs/FLUX.1-dev built for SGLang Diffusion.

What is inside

• —ModelOpt fp4 export of the FLUX.1-dev transformer
• —SGLang mixed-transformer post-processing for validated BF16 fallback layers
• —swapweightnibbles set to false in the quantization config for this validated export family

Intended usage

• —Keep the base FLUX.1-dev model separate
• —Download this transformer override locally
• —Load it with the transformer-path flag in SGLang

Example sglang generate --model-path black-forest-labs/FLUX.1-dev --transformer-path /path/to/this-repo --prompt-path /tmp/prompt.txt --num-gpus 4 --enable-torch-compile false

This is not a full standalone FLUX.1-dev model repo.