Help with OmniGen Pipeline

[hitchhiker3010]

Hi @a-r-r-o-w and team

I was trying out the OmniGen model pipeline it works great but the kwargs [use_kv_cache=True,offload_kv_cache=True,separate_cfg_infer=True] that were supported in the original model repo [https://github.com/VectorSpaceLab/OmniGen/blob/main/docs/inference.md] don't seem to work when used in the diffusers pipeline. Any direction on how this can be used will be great help. I went through the HF dicourse, HF discord and didn't find anything.
I was directed to raise an issue here so here I am.
Would love to understand

1. If there is any other material that I could use to improve the inference speed of OmniGen Pipeline.
2. How I could load the quantized models on to the pipeline. Example https://huggingface.co/gryan/OmniGen-v1-bnb-4bit
3. How to use kwargs like use_kv_cache=True,offload_kv_cache=True,separate_cfg_infer=True with the diffusers OmniGenPipeline.

Thanks in advance