ValueError: `position_ids should be a 3D tensor of shape (4, batch_size, seq_length)` when using agent-lightning 0.2.1 + Qwen2.5-VL

[WGowi]

Hello — I am currently working on replicating the “DeepEyes”-style pipeline using agent-lightning (v0.2.1) together with Qwen2.5-VL . When I run the model I encounter the following error:

ValueError: position_ids should be a 3D tensor of shape (4, batch_size, seq_length).

• agent-lightning version: 0.2.1
• verl version: 0.6.0

[ultmaster]

Qwen2.5VL is known to be problematic with current VERL integration. The major problem is we didn't upgrade our VERL integration after VERL has upgraded themselves.

Related to #105