FLUX.2-klein-4B/9B not supported in custom models (text encoder mismatch) #4591
Description
Feature request / 功能建议
Requested support for black-forest-labs/FLUX.2-klein-9B and black-forest-labs/FLUX.2-klein-4B
FLUX.2-klein uses Qwen3 as text encoder, but Xinference's FLUX.2 integration
(2.0.0) only supports Mistral3 (from FLUX.2-dev). When attempting to load
FLUX.2-klein as a custom model, Diffusers validation fails due to text encoder
type mismatch, leaving the pipeline with meta tensors.
Error: Expected types for text_encoder: Mistral3..., got Qwen3...
Then: Cannot copy out of meta tensor; no data!
Please add explicit support for FLUX.2-klein's Qwen3 text encoder.
Motivation / 动机
Smaller models are very useful for smaller GPU cards
Your contribution / 您的贡献
https://huggingface.co/black-forest-labs/FLUX.2-klein-9B#using-with-diffusers-🧨