support

[ayttop]

pipe = OmniGenPipeline.from_pretrained("Shitao/OmniGen-v1")

not support

device_map=auto.....acclrate

4-bit........bitsandbytes

[staoxiao]

Current code doesn't support quantization. We will consider this in the future.

[nitinmukesh]

@ayttop

If you want to run on low VRAM

try this
https://www.youtube.com/watch?v=9ZXmXA2AJZ4

[ayttop]

It never works in Colab T4 and I think it's because of the 12GB RAM. Is it possible to modify the model to 8 bit and upload it independently?

[ayttop]

How can I compress the form without using it in Gradio? I mean the compression command from the form path and then saving it compressed to a folder so that I can upload it to Hagface

[ayttop]

Manni1000
Can the form be compressed and uploaded compressed?
8-bit Quantization

[ayttop]

https://huggingface.co/goodasdgood/OmniGen_quantization/tree/main

[ayttop]

https://huggingface.co/goodasdgood/OmniGen_quantization/tree/main

not run

[werruww]

from OmniGen import OmniGenPipeline

pipe = OmniGenPipeline.from_pretrained("Shitao/OmniGen-v1", Quantization=True)

Text to Image

images = pipe(
prompt="A curly-haired man in a red shirt is drinking tea.",
height=1024,
width=1024,
guidance_scale=2.5,
num_inference_steps=2, # Changed argument name to num_inference_steps
seed=0,
)
images[0].save("example_t2i.png") # save output PIL Image

[werruww]

[werruww]

colab tpu

[werruww]

from OmniGen import OmniGenPipeline
import torch

pipe = OmniGenPipeline.from_pretrained("Shitao/OmniGen-v1", Quantization=True)

[werruww]

how to save OmniGen-v1 Quantization to folder in 8bit?

[werruww]

import torch

تحميل النموذج مع التكميم

model = OmniGen.from_pretrained("Shitao/OmniGen-v1", quantize=True)

حفظ الأوزان إلى ملف

torch.save(model.state_dict(), "/content/1/quantized_model.pth")

print("Model weights saved successfully.")

[werruww]

https://huggingface.co/sdyy/my-new-model_OmniGen-v1

[werruww]

from OmniGen import OmniGenPipeline

pipe = OmniGenPipeline.from_pretrained("sdyy/my-new-model_OmniGen-v1", Quantization=True)

pipe.save_pretrained("/content/a")