Repository navigation
CI failure due to transformers VLM change #723
Description
🐛 Describe the bug
CI details: Qwen2VLConfig, monkey patch impl related
Text config is seperated out from the general config in transformers>=4.52.0.
Qwen2VLRotaryEmbedding takes Qwen2VLTextConfig intead of Qwen2VLConfig now.
huggingface/transformers#37268
We also need to check all VLMs monkey patch since most models are refactored into TextModel and VisionModel
blocked on: huggingface/transformers#38331 (can walkaround by passing hidden_size and vocab_size to config dict directly though) fixed by huggingface/transformers#38372
Reproduce
No response
Versions
liger_kernel==0.5.9
transformers>=4.52.0
Activity
[-]CI failure due to Qwen2VLConfig change[/-][+]CI failure due to transformers VLM change[/+]Tcc0403 commentedon May 26, 2025
video_processor(Qwen2VLVideoProcessor) is now a required input forQwen2VLProcessor.We might need to add torchvision to our dev dependency since it is required for videoprocessor.
Related PR: huggingface/transformers#35206
fix: Fix llava (#743)
build(dev-dependency): Add torchvision for multimodal test (#755)
Tcc0403 commentedon Jun 13, 2025
fixed by #735 #738 #743 #755