阅读量:2
在近两年内，有关 LLM 的研究进展很快，每天几乎都有新的语言模型发布（隐藏的 GPT-5，Llama3，Qwen1.5，Mixtral 8x22B 和 Claude 3 等等等等），它们的性能和效果似乎每天都在持续提升。
然而，令人震惊的是，大多数现代 LLM 所使用的架构与最初的 GPT 模型非常相似。从模型架构角度出发，LLM 的一个关键组成部分一直保持不变，那就是 Transformer 架构的 Decoder。所有人或机构几乎都只是在将模型做得更大，对结构稍作修改，使用更大规模和更高质量的数据集，并采用更加先进的训练（和对齐）方法训练模型。
因此，深入了解 LLM 的内部结构对于研究人员和技术开发者来说是至关重要的。这不仅有助于我们更好地理解模型的性能和局限性，还能够指导我们如何更有效地设计和优化未来的模型。接下来，我将会简要地概述 LLM 常用的架构配置，然后针对国内外 30 个热门大模型的架构，进行详细的图文分析，以便大家对大模型有更深刻的理解。
架构配置
Transformer
架构类型
大型语言模型（LLMs）主要分为自回归模型、自编码模型和序列到序列模型这三种类型。这些模型几乎普遍采用 Transformer 架构，而传统的 RNN 架构则较少使用。基于 Transformer 架构的 LLMs，根据其设计特点，主要可以分为三类：仅包含编码器（Encoder-only）的模型、仅包含解码器（Decoder-only）的模型，以及同时包含编码器和解码器（Encoder-Decoder）的模型。
1. RNN
2. 基于 Transformer 的 GPT，BERT 和 Transformer XL
混合专家模型 MoE
MoE（Mixture-of-Experts）是一种神经网络架构，它通过路由机制（Router）将输入数据动态地分配给多个专家（Expert）网络中的一组。这种架构允许模型根据输入数据的特性选择不同的专家来处理，从而提高了模型的表达能力和效率。
1. 稀疏 MoE
2. 细粒度 MoE
注意力机制
注意力机制是一种在大语言模型中模拟人类注意力的技术，它通过动态调整输入数据的权重，使模型能够集中处理信息中最关键的部分。
1. 多头注意力
2. 稀疏注意力
3. 滑动窗口注意力
1MHA，GQA 和 MQA
位置编码
在 LLM 中，位置编码是一种将序列中 Token 的位置信息编码为模型可以理解和利用的方式的技术。
1. 基于正弦函数和余弦函数的固定位置编码
2. 可学习的（learnable）位置编码
3. ALiBi 位置编码
4. RoPE 位置编码
归一化
在 LLM 中，归一化是一种数据处理技术，通过将输入特征缩放到统一的尺度上，来提高模型的泛化能力和训练效率。
- Pre-Norm 和 Post-Norm
2. Pre-Norm
Sublayer 表示自注意力层或前馈神经网络层。
3. Post-Norm
4. LayerNorm
5. RMSNorm
RMSNorm 省略了 LayerNorm 中平均值μ的计算，只基于均方根进行缩放。
激活函数
激活函数是神经网络中的一种函数，用于对输入信号进行非线性变换，增加网络的表达能力。激活函数的选择对神经网络的性能和训练速度有很大的影响。
1. GeLU 和 SiLU
2. GLU( Gated Linear Units)
3. GeGLU 和 SwiGLU
详细架构
BERT
BERT 模型建立在 Transformer 的 Encoder 的基础上。
1. 模型架构
BertModel( (embeddings): BertEmbeddings( (word_embeddings): Embedding(28996, 768, padding_idx=0) (position_embeddings): Embedding(512, 768) (token_type_embeddings): Embedding(2, 768) (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True) (dropout): Dropout(p=0.1, inplace=False) ) (encoder): BertEncoder( (layer): ModuleList( (0-11): 12 x BertLayer( (attention): BertAttention( (self): BertSelfAttention( (query): Linear(in_features=768, out_features=768, bias=True) (key): Linear(in_features=768, out_features=768, bias=True) (value): Linear(in_features=768, out_features=768, bias=True) (dropout): Dropout(p=0.1, inplace=False) ) (output): BertSelfOutput( (dense): Linear(in_features=768, out_features=768, bias=True) (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True) (dropout): Dropout(p=0.1, inplace=False) ) ) (intermediate): BertIntermediate( (dense): Linear(in_features=768, out_features=3072, bias=True) (intermediate_act_fn): GELUActivation() ) (output): BertOutput( (dense): Linear(in_features=3072, out_features=768, bias=True) (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True) (dropout): Dropout(p=0.1, inplace=False) ) ) ) ) (pooler): BertPooler( (dense): Linear(in_features=768, out_features=768, bias=True) (activation): Tanh() ) ) 2. 图示
ChatGLM, ChatGLM2, ChatGLM3, GLM
GLM 是一个基于自动回归式空白填充预训练的通用语言模型。ChatGLM2 和 ChatGLM3 对 ChatGLM 的主要改进在于使用了 MAQ 注意力机制，RMSNorm 归一化方法和 SwiGLU 激活函数。下面仅给出 ChatGLM2 的模型架构。
1. 模型架构
ChatGLMForConditionalGeneration( (transformer): ChatGLMModel( (embedding): Embedding( (word_embeddings): Embedding(65024, 4096) ) (rotary_pos_emb): RotaryEmbedding() (encoder): GLMTransformer( (layers): ModuleList( (0-27): 28 x GLMBlock( (input_layernorm): RMSNorm() (self_attention): SelfAttention( (query_key_value): Linear(in_features=4096, out_features=4608, bias=True) (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) ) ) 2. 图示
ChatRWKV
RWKV 是一种基于 RNN 架构，并结合 Transformer 的优势的语言模型。
1. 模型架构图示
Command-R
Command-R 基于 Transformer 的 Decoder 进行创新和改进，具备 RAG（Retrieval Augmented Generation）的功能特性，部分模型使用了 GQA 技术。
1. 模型结构
CohereForCausalLM( (model): CohereModel( (embed_tokens): Embedding(256000, 8192, padding_idx=0) (layers): ModuleList( (0-39): 40 x CohereDecoderLayer( (self_attn): CohereAttention( (q_proj): Linear(in_features=8192, out_features=8192, bias=False) (k_proj): Linear(in_features=8192, out_features=8192, bias=False) (v_proj): Linear(in_features=8192, out_features=8192, bias=False) (o_proj): Linear(in_features=8192, out_features=8192, bias=False) (rotary_emb): CohereRotaryEmbedding() ) (mlp): CohereMLP( (gate_proj): Linear(in_features=8192, out_features=22528, bias=False) (up_proj): Linear(in_features=8192, out_features=22528, bias=False) (down_proj): Linear(in_features=22528, out_features=8192, bias=False) (act_fn): SiLU() ) (input_layernorm): CohereLayerNorm() ) ) (norm): CohereLayerNorm() ) (lm_head): Linear(in_features=8192, out_features=256000, bias=False) ) 2. 图示
GPT
GPT 模型建立在 Transformer 的 Decoder 的基础上。
1. 模型架构
`OpenAIGPTLMHeadModel( (transformer): OpenAIGPTModel( (tokens_embed): Embedding(40478, 768) (positions_embed): Embedding(512, 768) (drop): Dropout(p=0.1, inplace=False) (h): ModuleList( (0-11): 12 x Block( (attn): Attention( (c_attn): Conv1D() (c_proj): Conv1D() (attn_dropout): Dropout(p=0.1, inplace=False) (resid_dropout): Dropout(p=0.1, inplace=False) ) (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True) (mlp): MLP( (c_fc): Conv1D() (c_proj): Conv1D() (act): NewGELUActivation() (dropout): Dropout(p=0.1, inplace=False) ) (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True) ) ) ) (lm_head): Linear(in_features=768, out_features=40478, bias=False) )` 2. 图示
GPT2, GPT3, Falcon
GPT3 与 GPT2 在模型架构上的差别在于前者使用了稀疏注意力模式的注意力机制，而 Falcon 在 GPT3 上进行的最大变动在于前者使用了 RoPE 和 MQA，这里仅提供 GPT2 的模型架构，详情请查看对应模型的具体实现。
1. 模型架构
GPT2LMHeadModel( (transformer): GPT2Model( (wte): Embedding(50257, 768) (wpe): Embedding(1024, 768) (drop): Dropout(p=0.1, inplace=False) (h): ModuleList( (0-11): 12 x GPT2Block( (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True) (attn): GPT2Attention( (c_attn): Conv1D() (c_proj): Conv1D() (attn_dropout): Dropout(p=0.1, inplace=False) (resid_dropout): Dropout(p=0.1, inplace=False) ) (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True) (mlp): GPT2MLP( (c_fc): Conv1D() (c_proj): Conv1D() (act): NewGELUActivation() (dropout): Dropout(p=0.1, inplace=False) ) ) ) (ln_f): LayerNorm((768,), eps=1e-05, elementwise_affine=True) ) (lm_head): Linear(in_features=768, out_features=50257, bias=False) ) 2. 图示
Gemma
Gemma 在 Transformer 的 Decoder 上进行创新和改进，使用了 MQA 技术。
1. 模型架构
`GemmaForCausalLM( (model): GemmaModel( (embed_tokens): Embedding(256000, 2048, padding_idx=0) (layers): ModuleList( (0-17): 18 x GemmaDecoderLayer( (self_attn): GemmaSdpaAttention( (q_proj): Linear(in_features=2048, out_features=2048, bias=False) (k_proj): Linear(in_features=2048, out_features=256, bias=False) (v_proj): Linear(in_features=2048, out_features=256, bias=False) (o_proj): Linear(in_features=2048, out_features=2048, bias=False) (rotary_emb): GemmaRotaryEmbedding() ) (mlp): GemmaMLP( (gate_proj): Linear(in_features=2048, out_features=16384, bias=False) (up_proj): Linear(in_features=2048, out_features=16384, bias=False) (down_proj): Linear(in_features=16384, out_features=2048, bias=False) (act_fn): PytorchGELUTanh() ) (input_layernorm): GemmaRMSNorm() (post_attention_layernorm): GemmaRMSNorm() ) ) (norm): GemmaRMSNorm() ) (lm_head): Linear(in_features=2048, out_features=256000, bias=False) )` 2. 图示
Grok-1
Grok-1 在 Transformer 的 Decoder 上进行创新和改进，使用了 GQA ， MoE 和 Sandwish Normalization 等技术。
1. 模型架构
`Grok1ModelForCausalLM( (model): Grok1Model( (embed_tokens): Embedding(131072, 6144, padding_idx=0) (layers): ModuleList( (0-63): 64 x DecoderLayer( (attn): MultiHeadAttention( (q_proj): Linear(in_features=6144, out_features=6144, bias=False) (k_proj): Linear(in_features=6144, out_features=6144, bias=False) (v_proj): Linear(in_features=6144, out_features=6144, bias=False) (o_proj): Linear(in_features=6144, out_features=6144, bias=False) (rotary_emb): RotaryEmbedding() ) (moe_block): MoeBlock( (gate): Linear(in_features=6144, out_features=8, bias=False) (experts): ModuleList( (0-7): 8 x MoeMLP( (linear_v): Linear(in_features=6144, out_features=32768, bias=False) (linear_1): Linear(in_features=32768, out_features=6144, bias=False) (linear): Linear(in_features=6144, out_features=32768, bias=False) (act_fn): GELU(approximate='none') ) ) ) (pre_attn_norm): RMSNorm() (post_attn_norm): RMSNorm() (pre_moe_norm): RMSNorm() (post_moe_norm): RMSNorm() ) ) (norm): RMSNorm() ) (lm_head): Linear(in_features=6144, out_features=131072, bias=False) )` 2. 图示
LLama, LLama2, LLama3, Baichuan, Baichuan2, DeepSeek, DeepSeek-Coder, Intern, Intern2, OLMo,Phi-3, Skywork, Yi
Llama 基于 Transformer 的 Decoder 进行创新和改进，并吸收了 GPT3 和 PaLM 等最新研究的优点。
从 Baichuan2，Intern2，DeepSeek，OLMo，Phi-3，Skywork 和 Yi 的论文或技术报告了解到，其采用了和 LLama 相似的模型结构设计，但注意，实际上 Baichuan2，Intern2，DeepSeek，OLMo，Phi-3，Skywork 和 Yi 与 Llama2 并不完全一样，在注意力机制，位置编码，归一化，前馈神经网络的处理上可能存在着细微的差别，详情请查看对应模型的具体实现。而由于 LLama，LLama2 和 LLama3 的模型结构变化不大，这里给出 LLama2 的模型结构。
1. 模型架构
`LlamaForCausalLM( (model): LlamaModel( (embed_tokens): Embedding(32000, 4096) (layers): ModuleList( (0-31): 32 x LlamaDecoderLayer( (self_attn): LlamaAttention( (q_proj): Linear(in_features=4096, out_features=4096, bias=False) (k_proj): Linear(in_features=4096, out_features=4096, bias=False) (v_proj): Linear(in_features=4096, out_features=4096, bias=False) (o_proj): Linear(in_features=4096, out_features=4096, bias=False) (rotary_emb): LlamaRotaryEmbedding() ) (mlp): LlamaMLP( (gate_proj): Linear(in_features=4096, out_features=11008, bias=False) (up_proj): Linear(in_features=4096, out_features=11008, bias=False) (down_proj): Linear(in_features=11008, out_features=4096, bias=False) (act_fn): SiLUActivation() ) (input_layernorm): LlamaRMSNorm() (post_attention_layernorm): LlamaRMSNorm() ) ) (norm): LlamaRMSNorm() ) (lm_head): Linear(in_features=4096, out_features=32000, bias=False) )` 2. 图示
Mistral
Mistral 基于 Transformer 的 Decoder 进行创新和改进，使用了 GQA 和 SWA 等技术，部分模型采用了 MoE 架构。
1. 模型架构
MistralForCausalLM( (model): MistralModel( (embed_tokens): Embedding(32000, 4096) (layers): ModuleList( (0-31): 32 x MistralDecoderLayer( (self_attn): MistralAttention( (q_proj): Linear(in_features=4096, out_features=4096, bias=False) (k_proj): Linear(in_features=4096, out_features=1024, bias=False) (v_proj): Linear(in_features=4096, out_features=1024, bias=False) (o_proj): Linear(in_features=4096, out_features=4096, bias=False) (rotary_emb): MistralRotaryEmbedding() ) (mlp): MistralMLP( (gate_proj): Linear(in_features=4096, out_features=14336, bias=False) (up_proj): Linear(in_features=4096, out_features=14336, bias=False) (down_proj): Linear(in_features=14336, out_features=4096, bias=False) (act_fn): SiLU() ) (input_layernorm): MistralRMSNorm() (post_attention_layernorm): MistralRMSNorm() ) ) (norm): MistralRMSNorm() ) (lm_head): Linear(in_features=4096, out_features=32000, bias=False) ) 2. 图示
OpenELM
OpenELM 基于 Transformer 的 Decoder 进行创新和改进，使用了 GQA 和 Layer-wise scaling 等技术。
1. 模型架构
`OpenELMForCausalLM( (transformer): OpenELMModel( (token_embeddings): Embedding(32000, 1280) (layers): ModuleList( (0): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=12, key_heads=3, value_heads=3 (qkv_proj): Linear(in_features=1280, out_features=1152, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=768, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=1536, bias=False) (proj_2): Linear(in_features=768, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (1): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=12, key_heads=3, value_heads=3 (qkv_proj): Linear(in_features=1280, out_features=1152, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=768, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=2048, bias=False) (proj_2): Linear(in_features=1024, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (2): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=12, key_heads=3, value_heads=3 (qkv_proj): Linear(in_features=1280, out_features=1152, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=768, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=2560, bias=False) (proj_2): Linear(in_features=1280, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (3): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=12, key_heads=3, value_heads=3 (qkv_proj): Linear(in_features=1280, out_features=1152, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=768, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=3072, bias=False) (proj_2): Linear(in_features=1536, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (4): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=12, key_heads=3, value_heads=3 (qkv_proj): Linear(in_features=1280, out_features=1152, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=768, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=3584, bias=False) (proj_2): Linear(in_features=1792, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (5): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=4096, bias=False) (proj_2): Linear(in_features=2048, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (6): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=5120, bias=False) (proj_2): Linear(in_features=2560, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (7): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=5632, bias=False) (proj_2): Linear(in_features=2816, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (8): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=6144, bias=False) (proj_2): Linear(in_features=3072, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (9): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=6656, bias=False) (proj_2): Linear(in_features=3328, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (10): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=7168, bias=False) (proj_2): Linear(in_features=3584, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (11): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=16, key_heads=4, value_heads=4 (qkv_proj): Linear(in_features=1280, out_features=1536, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1024, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=7680, bias=False) (proj_2): Linear(in_features=3840, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (12): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=20, key_heads=5, value_heads=5 (qkv_proj): Linear(in_features=1280, out_features=1920, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1280, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=8704, bias=False) (proj_2): Linear(in_features=4352, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (13): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=20, key_heads=5, value_heads=5 (qkv_proj): Linear(in_features=1280, out_features=1920, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1280, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=9216, bias=False) (proj_2): Linear(in_features=4608, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (14): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=20, key_heads=5, value_heads=5 (qkv_proj): Linear(in_features=1280, out_features=1920, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1280, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=9728, bias=False) (proj_2): Linear(in_features=4864, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) (15): OpenELMDecoderLayer( (attn): OpenELMMultiHeadCausalAttention( query_heads=20, key_heads=5, value_heads=5 (qkv_proj): Linear(in_features=1280, out_features=1920, bias=False) (pos_embedding): OpenELMRotaryEmbedding( model_dim=64, max_seq_length=4096, freq_constant=10000) (q_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (k_norm): OpenELMRMSNorm(num_features=64, eps=1e-06) (out_proj): Linear(in_features=1280, out_features=1280, bias=False) ) (ffn): OpenELMFeedForwardNetwork( (ffn_with_glu) : True (proj_1): Linear(in_features=1280, out_features=10240, bias=False) (proj_2): Linear(in_features=5120, out_features=1280, bias=False) (act): SiLU() ) (ffn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) (attn_norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) ) (norm): OpenELMRMSNorm(num_features=1280, eps=1e-06) ) )` 2. 图示
Qwen, Qwen1.5
Qwen 基于 Transformer 的 Decoder 进行改进的语言模型，吸收了 Llama 的优点，而 Qwen1.5（Qwen2 的 beta 版本）进一步地在 Qwen 的基础上进行创新，并借鉴了 Mistral 的成功经验。Qwen1.5 对 Qwen 主要改进在于使用了 GQA，SWA（Longformer 和 Sparse Transformers），部分模型采用了 MoE 架构，更多细节请查看具体实现。下面给出 Qwen1.5/Qwen2 的模型结构及其图示。
1. 模型架构
`# Qwen2 Qwen2ForCausalLM( (model): Qwen2Model( (embed_tokens): Embedding(151936, 1024) (layers): ModuleList( (0-23): 24 x Qwen2DecoderLayer( (self_attn): Qwen2SdpaAttention( (q_proj): Linear(in_features=1024, out_features=1024, bias=True) (k_proj): Linear(in_features=1024, out_features=1024, bias=True) (v_proj): Linear(in_features=1024, out_features=1024, bias=True) (o_proj): Linear(in_features=1024, out_features=1024, bias=False) (rotary_emb): Qwen2RotaryEmbedding() ) (mlp): Qwen2MLP( (gate_proj): Linear(in_features=1024, out_features=2816, bias=False) (up_proj): Linear(in_features=1024, out_features=2816, bias=False) (down_proj): Linear(in_features=2816, out_features=1024, bias=False) (act_fn): SiLU() ) (input_layernorm): Qwen2RMSNorm() (post_attention_layernorm): Qwen2RMSNorm() ) ) (norm): Qwen2RMSNorm() ) (lm_head): Linear(in_features=1024, out_features=151936, bias=False) ) # Qwen2Moe Qwen2MoeForCausalLM( (model): Qwen2MoeModel( (embed_tokens): Embedding(151936, 2048) (layers): ModuleList( (0-23): 24 x Qwen2MoeDecoderLayer( (self_attn): Qwen2MoeAttention( (q_proj): Linear(in_features=2048, out_features=2048, bias=True) (k_proj): Linear(in_features=2048, out_features=2048, bias=True) (v_proj): Linear(in_features=2048, out_features=2048, bias=True) (o_proj): Linear(in_features=2048, out_features=2048, bias=False) (rotary_emb): Qwen2MoeRotaryEmbedding() ) (mlp): Qwen2MoeSparseMoeBlock( (gate): Linear(in_features=2048, out_features=60, bias=False) (experts): ModuleList( (0-59): 60 x Qwen2MoeMLP( (gate_proj): Linear(in_features=2048, out_features=1408, bias=False) (up_proj): Linear(in_features=2048, out_features=1408, bias=False) (down_proj): Linear(in_features=1408, out_features=2048, bias=False) (act_fn): SiLU() ) ) (shared_expert): Qwen2MoeMLP( (gate_proj): Linear(in_features=2048, out_features=5632, bias=False) (up_proj): Linear(in_features=2048, out_features=5632, bias=False) (down_proj): Linear(in_features=5632, out_features=2048, bias=False) (act_fn): SiLU() ) (shared_expert_gate): Linear(in_features=2048, out_features=1, bias=False) ) (input_layernorm): Qwen2MoeRMSNorm() (post_attention_layernorm): Qwen2MoeRMSNorm() ) ) (norm): Qwen2MoeRMSNorm() ) (lm_head): Linear(in_features=2048, out_features=151936, bias=False) )` 2. 图示
T5
T5 是一种与原始 Transformer 非常相似的 encoder-decoder 架构的模型。
1. 模型架构
T5ForConditionalGeneration( (shared): Embedding(32128, 768) (encoder): T5Stack( (embed_tokens): Embedding(32128, 768) (block): ModuleList( (0): T5Block( (layer): ModuleList( (0): T5LayerSelfAttention( (SelfAttention): T5Attention( (q): Linear(in_features=768, out_features=768, bias=False) (k): Linear(in_features=768, out_features=768, bias=False) (v): Linear(in_features=768, out_features=768, bias=False) (o): Linear(in_features=768, out_features=768, bias=False) (relative_attention_bias): Embedding(32, 12) ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (1): T5LayerFF( (DenseReluDense): T5DenseActDense( (wi): Linear(in_features=768, out_features=3072, bias=False) (wo): Linear(in_features=3072, out_features=768, bias=False) (dropout): Dropout(p=0.1, inplace=False) (act): ReLU() ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) ) ) (1-11): 11 x T5Block( (layer): ModuleList( (0): T5LayerSelfAttention( (SelfAttention): T5Attention( (q): Linear(in_features=768, out_features=768, bias=False) (k): Linear(in_features=768, out_features=768, bias=False) (v): Linear(in_features=768, out_features=768, bias=False) (o): Linear(in_features=768, out_features=768, bias=False) ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (1): T5LayerFF( (DenseReluDense): T5DenseActDense( (wi): Linear(in_features=768, out_features=3072, bias=False) (wo): Linear(in_features=3072, out_features=768, bias=False) (dropout): Dropout(p=0.1, inplace=False) (act): ReLU() ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) ) ) ) (final_layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (decoder): T5Stack( (embed_tokens): Embedding(32128, 768) (block): ModuleList( (0): T5Block( (layer): ModuleList( (0): T5LayerSelfAttention( (SelfAttention): T5Attention( (q): Linear(in_features=768, out_features=768, bias=False) (k): Linear(in_features=768, out_features=768, bias=False) (v): Linear(in_features=768, out_features=768, bias=False) (o): Linear(in_features=768, out_features=768, bias=False) (relative_attention_bias): Embedding(32, 12) ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (1): T5LayerCrossAttention( (EncDecAttention): T5Attention( (q): Linear(in_features=768, out_features=768, bias=False) (k): Linear(in_features=768, out_features=768, bias=False) (v): Linear(in_features=768, out_features=768, bias=False) (o): Linear(in_features=768, out_features=768, bias=False) ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (2): T5LayerFF( (DenseReluDense): T5DenseActDense( (wi): Linear(in_features=768, out_features=3072, bias=False) (wo): Linear(in_features=3072, out_features=768, bias=False) (dropout): Dropout(p=0.1, inplace=False) (act): ReLU() ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) ) ) (1-11): 11 x T5Block( (layer): ModuleList( (0): T5LayerSelfAttention( (SelfAttention): T5Attention( (q): Linear(in_features=768, out_features=768, bias=False) (k): Linear(in_features=768, out_features=768, bias=False) (v): Linear(in_features=768, out_features=768, bias=False) (o): Linear(in_features=768, out_features=768, bias=False) ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (1): T5LayerCrossAttention( (EncDecAttention): T5Attention( (q): Linear(in_features=768, out_features=768, bias=False) (k): Linear(in_features=768, out_features=768, bias=False) (v): Linear(in_features=768, out_features=768, bias=False) (o): Linear(in_features=768, out_features=768, bias=False) ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (2): T5LayerFF( (DenseReluDense): T5DenseActDense( (wi): Linear(in_features=768, out_features=3072, bias=False) (wo): Linear(in_features=3072, out_features=768, bias=False) (dropout): Dropout(p=0.1, inplace=False) (act): ReLU() ) (layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) ) ) ) (final_layer_norm): T5LayerNorm() (dropout): Dropout(p=0.1, inplace=False) ) (lm_head): Linear(in_features=768, out_features=32128, bias=False) ) 2. 图示
Yuan2
Yuan 在 Transformer 的 Decoder 进行改进，引入了一种新的注意力机制 LFA。
1. 模型架构
`YuanForCausalLM( (model): YuanModel( (embed_tokens): Embedding(135040, 2048, padding_idx=77185) (layers): ModuleList( (0-23): 24 x YuanDecoderLayer( (self_attn): YuanAttention( (v_proj): Linear(in_features=2048, out_features=2048, bias=False) (o_proj): Linear(in_features=2048, out_features=2048, bias=False) (rotary_emb): YuanRotaryEmbedding() (lf_gate): LocalizedFiltering( (conv1): Conv2d(2048, 1024, kernel_size=(2, 1), stride=(1, 1), padding=(1, 0)) (conv2): Conv2d(1024, 2048, kernel_size=(2, 1), stride=(1, 1), padding=(1, 0)) (output_layernorm): YuanRMSNorm() ) (q_proj): Linear(in_features=2048, out_features=2048, bias=False) (k_proj): Linear(in_features=2048, out_features=2048, bias=False) ) (mlp): YuanMLP( (up_proj): Linear(in_features=2048, out_features=8192, bias=False) (gate_proj): Linear(in_features=2048, out_features=8192, bias=False) (down_proj): Linear(in_features=8192, out_features=2048, bias=False) (act_fn): SiLU() ) (input_layernorm): YuanRMSNorm() (post_attention_layernorm): YuanRMSNorm() ) ) (norm): YuanRMSNorm() ) (lm_head): Linear(in_features=2048, out_features=135040, bias=False) )` 2. 图示
人工智能大模型越来越火了，离全民大模型的时代不远了，大模型应用场景非常多，不管是做主业还是副业或者别的都行，技多不压身，我这里有一份全套的大模型学习资料，希望给那些想学习大模型的小伙伴们一点帮助！
👉AI大模型学习路线汇总👈
大模型学习路线图，整体分为7个大的阶段：（全套教程文末领取哈）
第一阶段： 从大模型系统设计入手，讲解大模型的主要方法；
第二阶段： 在通过大模型提示词工程从Prompts角度入手更好发挥模型的作用；
第三阶段： 大模型平台应用开发借助阿里云PAI平台构建电商领域虚拟试衣系统；
第四阶段： 大模型知识库应用开发以LangChain框架为例，构建物流行业咨询智能问答系统；
第五阶段： 大模型微调开发借助以大健康、新零售、新媒体领域构建适合当前领域大模型；
第六阶段： 以SD多模态大模型为主，搭建了文生图小程序案例；
第七阶段： 以大模型平台应用与开发为主，通过星火大模型，文心大模型等成熟大模型构建大模型行业应用。
👉大模型实战案例👈
光学理论是没用的，要学会跟着一起做，要动手实操，才能将自己的所学运用到实际当中去，这时候可以搞点实战案例来学习。
👉大模型视频和PDF合集👈
观看零基础学习书籍和视频，看书籍和视频学习是最快捷也是最有效果的方式，跟着视频中老师的思路，从基础到深入，还是很容易入门的。
👉学会后的收获：👈
• 基于大模型全栈工程实现（前端、后端、产品经理、设计、数据分析等），通过这门课可获得不同能力；
• 能够利用大模型解决相关实际项目需求： 大数据时代，越来越多的企业和机构需要处理海量数据，利用大模型技术可以更好地处理这些数据，提高数据分析和决策的准确性。因此，掌握大模型应用开发技能，可以让程序员更好地应对实际项目需求；
• 基于大模型和企业数据AI应用开发，实现大模型理论、掌握GPU算力、硬件、LangChain开发框架和项目实战技能， 学会Fine-tuning垂直训练大模型（数据准备、数据蒸馏、大模型部署）一站式掌握；
• 能够完成时下热门大模型垂直领域模型训练能力，提高程序员的编码能力： 大模型应用开发需要掌握机器学习算法、深度学习框架等技术，这些技术的掌握可以提高程序员的编码能力和分析能力，让程序员更加熟练地编写高质量的代码。
👉获取方式：
😝有需要的小伙伴，可以保存图片到wx扫描二v码免费领取【保证100%免费】🆓
