internvl3.5 lora finetune code, would you release ?

[whcjb]

📚 The doc issue

is this be avaiable ?

Suggest a potential alternative/fix

No response

[Yuxin916]

I have the same question.

[ambermcheng-lang]

Same question

[wuzhaodongaipython]

请问您有找到全参微调脚本是哪个吗

[Yuxin916]

请问您有找到全参微调脚本是哪个吗

https://github.com/OpenGVLab/InternVL/tree/main/internvl_chat_gpt_oss/shell/internvl3_5_qwen3

sft

[wuzhaodongaipython]

请问您有找到全参微调脚本是哪个吗

https://github.com/OpenGVLab/InternVL/tree/main/internvl_chat_gpt_oss/shell/internvl3_5_qwen3

sft

谢谢

[Arseny5]

Hello! Guys, I made a clean, reproducible notebook for fine-tuning InternVL-3.5 on MVBench using multiple PEFT strategies (LoRA, QLoRA, and full-tuning). You can easily use it for fine-tuning on your tasks. Link: https://github.com/Arseny5/InternVL-3.5-QLoRA-Fine-tune

[xyn-abc]

[rank0]: AttributeError: 'InternVLChatModel' object has no attribute 'wrap_backbone_lora'

[siri-rouser]

Hi lads, I also meet the problem:AttributeError: 'InternVLChatModel' object has no attribute 'wrap_backbone_lora'

Here is what I did(helped by Gemini):

1. in file '''internvl_chat_finetune.py''', add '''from peft import LoraConfig, get_peft_model'''
2. replace: '''

if model_args.use_backbone_lora:

model.wrap_backbone_lora(r=model_args.use_backbone_lora, lora_alpha=2 * model_args.use_backbone_lora)

model.config.use_backbone_lora = model_args.use_backbone_lora

if model_args.use_llm_lora:

model.wrap_llm_lora(r=model_args.use_llm_lora, lora_alpha=2 * model_args.use_llm_lora)

model.config.use_llm_lora = model_args.use_llm_lora

''' to: ''' def manual_wrap_lora(model, args): """ Manually wraps the vision backbone and/or LLM with LoRA if the model class is missing the built-in methods. """ from peft import LoraConfig, get_peft_model

# 1. Wrap Vision Backbone
if args.use_backbone_lora > 0:
    print(f"Applying LoRA to Vision Backbone (r={args.use_backbone_lora})...")
    # Standard target modules for InternViT / ViT based backbones
    vision_target_modules = ['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2']
    
    vision_config = LoraConfig(
        r=args.use_backbone_lora,
        lora_alpha=2 * args.use_backbone_lora,
        target_modules=vision_target_modules,
        lora_dropout=0.05,
        bias='none',
    )
    # Apply PEFT specifically to the vision_model submodule
    model.vision_model = get_peft_model(model.vision_model, vision_config)
    model.vision_model.print_trainable_parameters()
    model.config.use_backbone_lora = args.use_backbone_lora

# 2. Wrap LLM
if args.use_llm_lora > 0:
    print(f"Applying LoRA to LLM (r={args.use_llm_lora})...")
    # Standard target modules for Llama/Qwen/InternLM style models
    llm_target_modules = ['q_proj', 'k_proj', 'v_proj', 'o_proj', 
                          'gate_proj', 'up_proj', 'down_proj']
    
    llm_config = LoraConfig(
        r=args.use_llm_lora,
        lora_alpha=2 * args.use_llm_lora,
        target_modules=llm_target_modules,
        lora_dropout=0.05,
        bias='none',
        task_type="CAUSAL_LM"
    )
    # Apply PEFT specifically to the language_model submodule
    model.language_model = get_peft_model(model.language_model, llm_config)
    model.language_model.print_trainable_parameters()
    model.config.use_llm_lora = args.use_llm_lora

return model

if hasattr(model, 'wrap_backbone_lora'): # If the model class supports it natively if model_args.use_backbone_lora: model.wrap_backbone_lora(r=model_args.use_backbone_lora, lora_alpha=2 * model_args.use_backbone_lora) model.config.use_backbone_lora = model_args.use_backbone_lora if model_args.use_llm_lora: model.wrap_llm_lora(r=model_args.use_llm_lora, lora_alpha=2 * model_args.use_llm_lora) model.config.use_llm_lora = model_args.use_llm_lora else: # Fallback to manual wrapper for InternVL 3.5 / Mismatched versions manual_wrap_lora(model, model_args) ''' 3. In '''internvl/model/internvl_chat/modeling_internvl_chat.py''' file , line 152 add this patch before for layer in self.language_model.model.layers:

''' llm_part = self.language_model.model if hasattr(llm_part, 'layers'): layers = llm_part.layers elif hasattr(llm_part, 'model') and hasattr(llm_part.model, 'layers'): layers = llm_part.model.layers else: raise AttributeError(f"Could not find layers in {type(llm_part)}")

for layer in layers: '''