对比各个大模型的网络结构

ps:使用自己的config,但是模型结构跟官方配置原理一致.

chatglm3

ChatGLMForConditionalGeneration(
  (transformer): ChatGLMModel(
    (embedding): Embedding(
      (word_embeddings): Embedding(65024, 4096)
    )
    (rotary_pos_emb): RotaryEmbedding()
    (encoder): GLMTransformer(
      (layers): ModuleList(
        (0-1): 2 x GLMBlock(
          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=12288, bias=False)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
        )
      )
      (final_layernorm): RMSNorm()
    )
    (output_layer): Linear(in_features=4096, out_features=65024, bias=False)
  )
)

#模型调试代码:
#=======通过下面这个代码可以debug学习chatglm3的代码.
from modeling_chatglm import ChatGLMForConditionalGeneration,ChatGLMConfig
import torch
 
def run ():
  config= ChatGLMConfig(num_layers=2,original_rope=True,use_cache=True)  #=====有一些参数在config.json里面搬过来即可.
  model = ChatGLMForConditionalGeneration(config=config)
  inputs_ids  = torch.randint(low=0,high=config.vocab_size, size=(4,30))
  print(model)
  res = model(inputs_ids)
  print(res)
 
run()

llama2

LlamaForCausalLM(
  (model): LlamaModel(
    (embed_tokens): Embedding(32000, 2048)
    (layers): ModuleList(
      (0-1): 2 x LlamaDecoderLayer(
        (self_attn): LlamaAttention(
          (q_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (k_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (v_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (o_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): LlamaMLP(
          (gate_proj): Linear(in_features=2048, out_features=554, bias=False)     
          (up_proj): Linear(in_features=2048, out_features=554, bias=False)       
          (down_proj): Linear(in_features=554, out_features=2048, bias=False)     
          (act_fn): SiLUActivation()
        )
        (input_layernorm): LlamaRMSNorm()
        (post_attention_layernorm): LlamaRMSNorm()
      )
    )
    (norm): LlamaRMSNorm()
  )
  (lm_head): Linear(in_features=2048, out_features=32000, bias=False)
)

#模型调试代码:
from transformers.models .llama import LlamaModel,LlamaConfig,LlamaForCausalLM
import torch
 
def run ():
  llamaconfig= LlamaConfig(vocab_size=32000,
  hidden_size=4096//2,
  intermediate_size=1108//2,
  num_hidden_layers=2,
  num_attention_heads=32//2,max_position_embeddings=2048//2)
  llamamodel = LlamaForCausalLM(config=llamaconfig) #https://hf-mirror.com/hiyouga/Llama-2-Chinese-13b-chat/blob/main/config.json 参考这里面的architecture知道llama2依然用的事llamaforcausallm架构.
  inputs_ids  = torch.randint(low=0,high=llamaconfig.vocab_size, size=(4,30))
  print(llamamodel)
  res = llamamodel(inputs_ids)
  print(res)
 
run()

核心部分放在一起比较

chatglm:          (input_layernorm): RMSNorm()
          (self_attention): SelfAttention(
            (query_key_value): Linear(in_features=4096, out_features=12288, bias=False)
            (core_attention): CoreAttention(
              (attention_dropout): Dropout(p=0.0, inplace=False)
            )
            (dense): Linear(in_features=4096, out_features=4096, bias=False)
          )
          (post_attention_layernorm): RMSNorm()
          (mlp): MLP(
            (dense_h_to_4h): Linear(in_features=4096, out_features=27392, bias=False)
            (dense_4h_to_h): Linear(in_features=13696, out_features=4096, bias=False)
          )
		  
		  
		  
		  
llama2:		(self_attn): LlamaAttention(
          (q_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (k_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (v_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (o_proj): Linear(in_features=2048, out_features=2048, bias=False)       
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): LlamaMLP(
          (gate_proj): Linear(in_features=2048, out_features=554, bias=False)     
          (up_proj): Linear(in_features=2048, out_features=554, bias=False)       
          (down_proj): Linear(in_features=554, out_features=2048, bias=False)     
          (act_fn): SiLUActivation()
        )
        (input_layernorm): LlamaRMSNorm()
        (post_attention_layernorm): LlamaRMSNorm()
		  
		  
		  
		  
		  
		  
		  

总结: 可以看出这2个模型除了参数有些区别外模型架构一模一样. chatglm的attendim=4096, dnn层 参数少. llama2的attentiondim=2048,dnn层参数多.就这么点区别.