Skip to content

[bug]: Flux.2 Klein 9B text encoder compatibility limitations #8839

Description

@NiceDragon8066

Is there an existing issue for this problem?

  • I have searched the existing issues

Install method

Invoke's Launcher

Operating system

Windows

GPU vendor

Nvidia (CUDA)

GPU model

GTX 1650

GPU VRAM

4GB

Version number

v6.11.0

Browser

Microsoft Edge

System Information

N/A

What happened

I have tried several different versions of the Qwen3-8B text encoder for the Flux.2 Klein 9B model in InvokeAI. Except for an FP16 version (approx 16 GB) all other versions of the text encoder failed to work in the InvokeAI App. The error messages are various, from tensor dimensions mismatch error to data type string when byte is expected error. I tried the Qwen3-8B text encoder InvokeAI tries to download (diffusers format, approx. 16GB), it failed with data type mismatch error, string vs byte. I tried to use quantized varieties such as fp8 and fp4, but those failed with tensor dimension mismatch errors. I also tried to use the currently working text encoder in quanto_bf16_int8 format from the Wan2GP App with the Klein 9B model, but that failed with unrecognized model type error. I was ready to try the Gemma3-8B encoder, but at that point the Qwen3-8B_fp16.safetensors (approx. 16GB) text encoder worked. I personally believe that the problem shows a bug or limitation of the InvokeAI App as it pertains to the Flux.2 Klein 9B model image generation. Note: I haven't faced this problem in any other App. I look forward to a fix in future version.

What you expected to happen

I expected the different varieties, especially the quantized ones, of the Qwen3-8B encoder to work to conserve memory. But that did not happen, only a specific version works and that too only with the Flux.2 Klein-9B_Q4_K_M.gguf version of the model. I have not been able get the fp8 version of the Klein 9B model to work at all so far. No problem with the Klein 4B model with any variety of the Qwen3-4B encoders. Please investigate and fix the issue if you find the cause.

How to reproduce the problem

Use a low VRAM computer with 32GB RAM, Windows 11, Klein 9B model (any size other than Q4 gguf), and any of the Qwen3-8B encoders out there including the ones InvokeAI is wired to download. Like I said above - only the fp16 variety works, that too only with the Q4_K_M gguf quantized Klein-9B model.

Additional context

No response

Discord username

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions