A ValueError is raised when training with the latest transformers version (v5.3.0) and liger-kernel for any of the following VL models:
glm4v
glm4v_moe
qwen2_vl
qwen2_5_vl
qwen3_vl
qwen3_vl_moe
The issue appeared after the merge of this PR in transformers:
I opened an upstream issue and PR to fix this in liger-kernel:
It is still pending review, approval, merge and release.
See previous related issue: