We keep a snapshot of the CogAgent repo here, which we use for finetuning and inference.
The code base is based on CogVLM/CogAgent and swissarmytransformer v0.4.10.
Please install necessary libraries following the instruction in CogVLM/CogAgent README.
The finetuning script is finetune_cogagent_lora_design2code.sh.
Note that the LoRA modules are only added to the language decoder.
We provide the example inference script and the Design2Code-18B-v0 model weight.