| # MemGen VIS Training - Math Vision (Single GPU) | |
| # Based on mm_math training configuration | |
| # Change to project directory | |
| cd /root/CVPR/MemGen | |
| # Set environment variables | |
| export DEBUG_MODE="true" | |
| export LOG_PATH="./test_output/debug_log_math_vision.txt" | |
| export CUDA_VISIBLE_DEVICES="0" | |
| # Run with single GPU | |
| uv run python main.py \ | |
| --cfg-path configs/latent_memory/math_vision.yaml \ | |
| --options \ | |
| model.reasoner_model_name "Qwen/Qwen2.5-VL-7B-Instruct" \ | |
| model.weaver.weaver_model_name "Qwen/Qwen2.5-1.5B-Instruct" \ | |
| model.trigger.trigger_model_name "null" \ | |
| model.weaver.prompt_latents_len "8" \ | |
| model.weaver.inference_latents_len "4" \ | |
| model.max_prompt_aug_num "0" \ | |
| model.max_inference_aug_num "3" \ | |
| model.load_model_path "null" \ | |
| run.mode "train" \ | |
| run.train_weaver "True" \ | |
| run.train_trigger "False" \ | |
| run.train_weaver_method "grpo" \ | |
| run.generation.do_sample "True" \ | |
| run.generation.temperature "1.0" \ | |
| run.generation.max_response_length "512" \ | |
| run.output_dir "/root/CVPR/MemGen/test_output/math_vision" \ | |
| datasets.math_vision.mode "grpo" | |