| from transformers import ( | |
| PretrainedConfig, | |
| Qwen2_5_VLForConditionalGeneration, | |
| # Qwen3VLForConditionalGeneration, | |
| ) | |
| from thinkstream.model.casia.casia_vl import Qwen3VLForConditionalGeneration | |
| MODEL_CLS = { | |
| "qwen2.5vl": Qwen2_5_VLForConditionalGeneration, | |
| "qwen3vl": Qwen3VLForConditionalGeneration, | |
| } | |
| DEFAULT_VIDEO_FLEX_WINDOW_SIZE = 20 | |
| def get_text_config(config: PretrainedConfig) -> PretrainedConfig: | |
| """Return the text backbone sub-config, handling both flat and nested layouts.""" | |
| tc = getattr(config, "text_config", None) | |
| return tc if tc is not None else config | |