Restore multimodality (Qwen3_5ForConditionalGeneration)

#3

Re-attaches the vision tower and MTP weights from Qwen/Qwen3.5-27B to restore full Qwen3_5ForConditionalGeneration compatibility.

Changes:

  • Architecture: Qwen3_5ForCausalLM β†’ Qwen3_5ForConditionalGeneration
  • Added vision tower weights (model.visual.*) from Qwen/Qwen3.5-27B
  • Added MTP weights (mtp.*) from Qwen/Qwen3.5-27B
  • Added preprocessor_config.json and video_preprocessor_config.json
  • Updated config.json with vision_config, image/video token IDs
  • Text weights are unchanged

LGTM!

ConicCat changed pull request status to merged

Sign up or log in to comment