add support and optimization for minicpmo vision part (#12713)
This commit is contained in:
		
							parent
							
								
									c52bdff76b
								
							
						
					
					
						commit
						b62734748f
					
				
					 1 changed files with 21 additions and 0 deletions
				
			
		| 
						 | 
				
			
			@ -1028,6 +1028,15 @@ def _optimize_pre(model, qtype=None):
 | 
			
		|||
            model.llm.config.model_type = "minicpm"
 | 
			
		||||
        _optimize_pre(model.llm, qtype=qtype)
 | 
			
		||||
        model.llm.config.model_type = "minicpmv"
 | 
			
		||||
    elif model.config.model_type == "minicpmo":
 | 
			
		||||
        # vpm opt
 | 
			
		||||
        from ipex_llm.transformers.models.minicpmv import merge_qkv
 | 
			
		||||
        model.vpm.apply(merge_qkv)
 | 
			
		||||
 | 
			
		||||
        # llm opt
 | 
			
		||||
        model.llm.config.model_type = "qwen2"
 | 
			
		||||
        _optimize_pre(model.llm, qtype=qtype)
 | 
			
		||||
        model.llm.config.model_type = "minicpmo"
 | 
			
		||||
    elif model.config.model_type == "megrezo":
 | 
			
		||||
        from ipex_llm.transformers.models.minicpmv import merge_qkv
 | 
			
		||||
        model.vision.apply(merge_qkv)
 | 
			
		||||
| 
						 | 
				
			
			@ -1944,6 +1953,18 @@ def _optimize_post(model):
 | 
			
		|||
            convert_forward(model.vpm, vpm_module.Idefics2VisionAttention, siglip_attention_forward)
 | 
			
		||||
            minicpmv_chat = minicpmv_chat_wrapper(module.MiniCPMV.chat)
 | 
			
		||||
            model.chat = MethodType(minicpmv_chat, model)
 | 
			
		||||
    elif model.config.model_type == "minicpmo":
 | 
			
		||||
        # vpm opt
 | 
			
		||||
        vpm_modeling_module_name = model.vpm.__class__.__module__
 | 
			
		||||
        vpm_module = importlib.import_module(vpm_modeling_module_name)
 | 
			
		||||
 | 
			
		||||
        from ipex_llm.transformers.models.minicpmv import siglip_attention_forward
 | 
			
		||||
        convert_forward(model.vpm, vpm_module.SiglipAttention, siglip_attention_forward)
 | 
			
		||||
 | 
			
		||||
        # llm opt
 | 
			
		||||
        model.llm.config.model_type = "qwen2"
 | 
			
		||||
        _optimize_post(model.llm)
 | 
			
		||||
        model.llm.config.model_type = "minicpmo"
 | 
			
		||||
    elif model.config.model_type == "megrezo":
 | 
			
		||||
        modeling_module_name = model.__class__.__module__
 | 
			
		||||
        module = importlib.import_module(modeling_module_name)
 | 
			
		||||
| 
						 | 
				
			
			
 | 
			
		|||
		Loading…
	
		Reference in a new issue