Your demo code on HuggingFace is throwing 502 Gateway error

[vbayanag]

Please provide instruction son how to evaluate AlphaCLIP MLLM model.

[SunzeY]

the openxlab resource is limited, the source code is available in demo, you can run it locally.

[vbayanag]

Hey SunzeY,
Your demo with_llm is throwing the following error. I'm running the code on rocm/pytorch :rocm6.1.2_ubuntu20.04_py3.9_pytorch_release-2.1.2 docker image on AMD GPU

ImportError: cannot import name '_expand_mask' from 'transformers.models.bloom.modeling_bloom' (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/transformers/models/bloom/modeling_bloom.py)

[SunzeY]

maybe the torchvision version is not matched? You should make sure to be able to run llava original codes first in your environment before testing alpha-clip. any environmental problem can be better solved in LLaVA repo.

[vbayanag]

Also, earlier I got an error at https://github.com/SunzeY/AlphaCLIP/blob/main/demo/with_llm/llava/model/language_model/llava_llama.py#L139

Llava is already used by transformers model.

Is it expected that I change the registration name on L#139 when I run the demo?

[SunzeY]

you can change transformers version to 4.36.2, as newer package already include llava in official model. If nothing went wrong after changing the name, it's also fine.

[vbayanag]

After changing the name I got the below error on transformers 4.42+
ImportError: cannot import name '_expand_mask' from 'transformers.models.bloom.modeling_bloom' (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/transformers/models/bloom/modeling_bloom.py)

Let me try with a different version now. Did you happen to get a chance to run a demo on AMD hardware? Otherwise, I'd be happy to help.

[SunzeY]

nop