Instructions to use LanguageBind/UniWorld-V1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- univa
How to use LanguageBind/UniWorld-V1 with univa:
# Follow installation instructions at https://github.com/PKU-YuanGroup/UniWorld-V1 from univa.models.qwen2p5vl.modeling_univa_qwen2p5vl import UnivaQwen2p5VLForConditionalGeneration model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained( "LanguageBind/UniWorld-V1", torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2", ).to("cuda") processor = AutoProcessor.from_pretrained("LanguageBind/UniWorld-V1") - Notebooks
- Google Colab
- Kaggle
Demo on HF Spaces
#5
by merve - opened
Hello! If you want to bring your demo to HF Spaces we'd love to assign a ZeroGPU grant :)
Sure, maybe I will do it in next few days. Thanks for your support.
Hello! If you want to bring your demo to HF Spaces we'd love to assign a ZeroGPU grant :)
@merve Hello! I’ve deployed UniWorld-V1 to a Space with zero GPU and my model is about 20 B. Even with @spaces.GPU(duration=600) it still times out during download and initialization. To work around this, I changed the code to download and initialize on CPU first (no GPU), then move the model to GPU in one go. However, CPU initialization now runs out of RAM (only 64 GB available). How can I fix this?
