test_configs/optimized_llama2_config.yaml

model:
  arch: mini_gpt4_llama_v2
  freeze_vit: True
  freeze_qformer: True
  max_txt_len: 256
  low_resource: True
  image_size: 224
  end_sym: "</s>"
  # 🔧 使用更小的模型以节省显存
  llama_model: "microsoft/Phi-3.5-mini-instruct"
  ckpt: "checkpoints/video_llama_checkpoint_last.pth"
  use_grad_checkpoint: True
  chat_template: True
  lora_r: 64
  lora_alpha: 16
  length: 50
  use_grad_checkpoint_llm: True
  max_context_len: 3600
  architectures: [
    "MiniGPT4_Video"
  ]
  device: "cuda"
  drop_path_rate: 0 
  img_size: 224
  model_type: "minigpt4_video"
  num_query_token: 32
  prompt: ""
  # 🔧 使用float32以避免精度问题
  torch_dtype: "float32"
  transformers_version: "4.42.3"
  vit_precision: "fp16"
  vit_model: "eva_clip_g"
  token_pooling: true
  lora_target_modules : ["q_proj","v_proj"]
  lora_dropout: 0.05
  remove_template: false
  prompt_path: ""
  minigpt4_gpu_id: 0
  whisper_gpu_id: 0
  answer_module_gpu_id: 0

datasets:
  video_chatgpt:
    batch_size: 2  # 🔧 减小batch size以节省显存
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 224
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 200

run:
  seed: 42
  amp: false  # 🔧 禁用AMP以避免兼容性问题